This PDF file contains the front matter associated with SPIE Proceedings Volume 12707, including the Title Page, Copyright information, Table of Contents, and Conference Committee listings.

The large blind region behind the huge body of a truck is the primary reason for reversing collisions. However, there are few truck reversing auxiliary approaches to avoid reversing accidents. In this paper, we present a novel heavy truck reversing safety approach based on pedestrian detection and tracking using binocular vision stitching. The proposed system contains three primary steps: binocular vision stitching, pedestrian detection and tracking, and heavy truck reversing speed control method. First of all, a binocular camera system based on vision stitching is used to perceive the reversing environment of the heavy truck. Secondly, the framework of YOLOv3 and discriminative correlation filter based tracking method is used to detect and track the pedestrian in real time. The pedestrian feature is extracted with improved MobileNetV2. Finally, a heavy truck reversing speed control method which can automatically control the reversing speed to avoid collisions improves the safety of truck reversing. This system has been tested on a van truck. Experiments demonstrate the viability of the proposed heavy truck reversing safety system.

In order to improve the ability to describe image information, this paper proposes an image retrieval method based on kmeans block. Firstly, under Lab color space, the image is divided into 6 sub-regions using k-means algorithm. And the external rectangle determined by each sub-region is segmented on the original image to obtain 6 sub-blocks that can contain the main information of the image. Then, the color features of each sub-block are extracted separately. Finally, the degree of similarity is reflected by the Euclidean distance between the image feature vectors. The simulation experimental results show that the proposed algorithm in this paper has been improved compared with the comparison scheme.

Image recognition technology, as an important branch in the field of artificial intelligence, aims to process and analyze images to recognize objects in images. The aim of this paper is to summarize the general approach to implementing image recognition and the application of this technology in some fields by analyzing relevant articles in the field of image recognition in recent years and to suggest possible conjectures on the future direction of this technology. This paper points out that the image recognition process is divided into two main parts: image processing and image recognition. For the implementation of the corresponding algorithms, this paper introduces two commonly used software libraries: OpenCV and YOLOv5. In addition, the paper also focuses on the applications and contributions of image recognition techniques to face recognition, lane detection, and medical image analysis. It is also pointed out that image recognition technology has a very promising future, and that there is still a lot of further improvement in the field of medical image analysis.

In order to solve the problem of high-speed image compression, this paper designs a parallel bit-plane encoding algorithm based on CCSDS image compression algorithm, and uses FPGA to implement it. Experimental results show that the system can effectively reduce the image coding time and meet the requirements of high-speed image compression.

In order to meet the demand of strong real-time display in large integrated video network, this paper proposes a lossless video strong real-time transmission display model based on fibre channel protocol. By accessing the original video data of multi-standard sensors, the normalized video format is packaged into fibre channel protocol data for long-distance lossless transmission and real-time display. Finally, the multi-format original video transmission display based on fibre channel FC-AE-ASM protocol is realized by FPGA accelerating platform, this technology can support the real-time display of multi-format and multi-resolution raw data, the maximum bandwidth is no less than 130MB/S, the maximum resolution is no less than the 1920 times 1080@60HZ, and the total system delay is no more than 0.79ms.

Infrared ship target recognition technology can automatically detect, analyze and identify ship targets, which is suitable for various types of working environments. This paper takes ship target image as the research object, and measures the effect of existing infrared detection technology, traditional target detection technology and target detection technology based on depth learning through data comparison. For infrared detection and target recognition, its working principle is summarized, and its ability to quickly identify ship types in practical applications is verified. The application and development status of Sea area planning、 supervision and Military reconnaissance field are summarized, and the future development trend is prospected.

Verifying individual Chinese handwritten signatures is an essential biometric technology that is widely used in banking, finance, and legal business. The forging of signatures for the purpose of cheating is a serious detriment to the interests of these industries. This paper proposes a Siamese network verification signature based on image domain transfer. The Siamese network uses a convolutional neural network as a sub-network to build the structure of Siamese network by combining the genuine signature network and the unauthenticated signature network. Each sub-network transfers ImageNet weights for training in the Chinese recognition task so that the new weighted image domain is suitable for the Chinese signature image domain. This Siamese network is trained with new weights to determine the authenticity of the signature. Currently, there is no publicly available dataset of Chinese handwritten signatures. This paper develops the Chinese signatures dataset from 295 induvial persons, 885 different persons participated, including approximately 9,000 pictures. The Siamese network achieves 92.75% accuracy on the test set, and the verification time for a single Chinese handwritten signature is 0.32 seconds. Finally, the Siamese network model achieves more than 90% accuracy on the public handwritten signature datasets CEDAR, BHSig-B and BHSig-H in three different languages, and the experiments demonstrate the good generalization of the proposed method.

Histobiology, development and regeneration, disease modeling, organ transplantation technology improvement, drug discovery/efficacy evaluation, and pathology research can be carried out through the parameters such as morphology, size, and number of organ like organs. However, it takes a lot of time to manually count various parameters of organs. Therefore, here we propose a new semantic segmentation model, AttUneXt, for segmentation of adenocarcinoid organs. The dice of the model can reach 0.9. As can be seen, our proposed model achieves the effect of rapidly and precisely segmenting organoids.

In recent years, due to the shortage of transmission line channel resources, power companies have begun to expand the capacity of existing important transmission lines. The safe and stable operation of important transmission lines is critical to the power supply status of users. Timely monitoring of potential hazards of construction machinery near the transmission lines has become a hot topic of transmission line protection against external damage. Aiming at the technical bottleneck of high false alarm rate of the current transmission line using online monitoring camera, this paper proposes a construction vehicle detection method based on the fusion of radar and visual features, which uses the physical features and geometric features of the target. The physical features such as velocity and acceleration are selected from radar. After the fusion of the radar data and camera data, the region of interest (ROI) of the radar target on the image is obtained, and the gradient direction histogram feature is extracted on the ROI. The visual features are calculated by the statistical features of gradient direction histogram, including standard deviation, median and average. This paper constructs a neural network R-V-DenseNet whose input is the fusion feature of radar and vision. Then a data set is made to train the network. The experimental results on the test set prove that the accuracy of R-V-DenseNet is improved compared with the traditional HOG-SVM method and the single sensor based detection method, which means the proposed method gains more accurate detection.

To solve the problem of insufficient representation of radar information in the fusion process, this paper proposes a fusion architecture based on attention enhancement. Considering that the projection area of radar point clouds can provide target’s position information, the spatial attention weight is obtained from radar images and then weighted to the camera images to supply the target’s position information from radar. At the same time, the transverse velocity, radial velocity and distance of the targets are assigned to each channel of the radar images, and the channel attention mechanism is used to enhance the channel feature of the radar images. We adopt RetinaNet with the backbone of Mobilenetv2 to build the fusion network architecture and perform target detection for radar images and camera images. The result shows that the proposed fusion method performs better than baseline network on nuScenes dataset, the mAP increased by 2.56% and the network is more lightweight.

As the application of panorama technology is increasing widely, improving the speed of 3D panorama construction and the quality of image stitching has become a popular research content nowadays. By improving the synchronization of multi-camera acquisition in panorama construction technology, the reliability of the stitched panorama images can be increased. At present, there are mainly two methods to ensure the synchronization of multi-camera acquisition. One is to ensure the synchronization of acquisition by using multi-camera acquisition cards through hardware, but the drawback of this method is the high requirement for the interface between cameras and the host computer; the other is to achieve synchronized acquisition through multi-thread control of the host computer through software, but the drawback of this method is the high synchronization delay. This paper introduces the polling system and multitasking system in hardware synchronization control, and experimentally explores the difference between the two systems for the delay in image acquisition control process and uses an MCU equipped with FreeRTOS real-time operating system to control multiple industrial cameras for multiple GPIO acquisition. The conclusion is that by using the FreeRTOS multitasking system for image acquisition control, the acquisition latency between multiple image acquisition devices can be greatly reduced.

With the construction of a large number of substations, the display problem of high-temperature equipment has been concerned. White light and infrared image registration and fusion method based on NSCT transform are designed to realize the display of high-temperature equipment in the substation. The white light and infrared images are enhanced based on NSCT. After the two images are calibrated, the white light and infrared images are registered by combining feature registration and edge detection methods. An image fusion algorithm is designed by combining wavelet transform and IHS transform to realize the fusion processing of two images. The test results indicate that the average gradient, information entropy, spatial frequency, and standard deviation of the designed method are higher. The image registration and fusion effect are in line with the human subjective vision, and the registration and fusion effect is better.

Previous works on 3D human pose estimation have concentrated on predicting the 3D pose of the human body from a single image, ignoring correlation between adjacent frames in video. We design a transformer network structure that can extract video temporal information, and enhance the accuracy of human pose prediction by encoding relative position with temporal fusion transformer structure to enhance local feature learning capability. On Human3.6M, we quantitatively and qualitatively analyze our method. Research suggests that our TSFormer achieves state-of-the-art performance.

According to the multi-layered system, environmental pollution monitoring and control information is developed through the use of network, database technology, GIS and other technologies, which improve the modern level of environmental monitoring and management. The system is characterized by technology, reasonable design, simple operation, and high intelligence.

Optical Coherence Tomography (OCT) uses low coherence light to provide a high spatial resolution to detect changes in the microstructure of living organisms in a non-invasive, real-time manner. A new OCT image denoising method is proposed to address the problem of poor noise reduction by conventional OCT image denoising algorithms. The method combines a Bayesian non-local mean filtering algorithm and deep learning to denoise noisy images for better noise reduction. By comparing with the Gaussian filtering algorithm, the median filtering algorithm, the BNLM (Bayesian nonlocal mean) denoising algorithm, the BM3D (block-matching and 3D filtering) denoising algorithm, the new algorithm outperforms the traditional method in terms of noise reduction. And compares the peak signal-to-noise ratio, structural similarity and other metrics. New algorithm shows superiority.

There are a lot of difficulties in the process of extracting function relationship from hard copy curve. Considering factors such as distortion, chromatic aberration, noise pollution and etc., a dense set of calibrated points near the real curve from the digitized image can be extracted out. Based on the fact that the function relationship corresponding to the curve can be expressed in the form of segmented Hermite polynomial function, a least squares model with outlier elimination strategy is presented, which can be solved to obtain the function parameters corresponding to the curve by iterative method. The simulation analysis shows that the difference between the obtained fitting function and the original function is positively related to the effective resolution of the digital image, and the algorithm provided in this paper is of high practical value.

Gaze tracking can be divided into two-dimensional (2-D) gaze-estimation method and three-dimensional (3-D) gazeestimation method. 3-D gaze-estimation method can simplify user calibration and track gaze under full free head movement. To solve the shortage of traditional 3-D gaze-estimation method, this paper proposes a 3-D eyeball optical axis (OA) reconstruction method for single-camera multi-source head-mounted system. The method directly calculates the vector of the OA from the pupil image on the basis of the corneal center, and then reconstructs the eyeball OA. The results of simulation experiments show the effectiveness of the method.

As a recognition method that does not require target cooperation, gait recognition has significant value in the criminal investigation field. As the gait recognition network’s input, the gait silhouette occupies a large storage space and low information density, which leads to a high demand for hardware in the gait recognition system. We proposed a compressing gait silhouette method based on the graphical edge, which extracts silhouette edges and combines them into a gait silhouette edge image. Storage space has been reduced by 90.6%, and information density has increased from 4.64% to 49.48%-71.09%. Since the new image is not applicable to the original network, we proposed the improved network GSEInet of GaitSet. The experimental results show that using our compressed images can reduce the network running time by 91.1%, and the demand for graphics card memory by 48.8%, while the recognition rate will not decrease. This will help transplant gait recognition tasks from large servers to small embedded devices and improve flexibility.

Supervised learning models require a large-scale dataset for more effective model training. However, a large-scale dataset requires time-consuming data collection and highly complicated preprocessing, posing a significant challenge for researchers. Therefore, creating a large-scale dataset for each practical application is unachievable in a real-world setting. In this case, unsupervised domain adaptation is crucial for practical applications, as it can enable models to enhance their performance on unlabeled data by training on labeled data. For multi-modal egocentric video analysis, some models have used unsupervised domain adaptation and achieved outstanding performance. However, they use either early or late fusion, ignoring the correlation in and between multi-modal inputs. Therefore, this paper investigates different fusion architecture and proposes a cascade attentional fusion to improve the feature representation for unsupervised domain adaptation on multi-modal egocentric video analysis. First, we propose a cascade fusion architecture for increased audio signal reuse. Then, we propose a temporal-spatial attention mechanism for highlighting spatio-temporal feature representations. Third, we propose a novel cascade attentional fusion method for multi-modal egocentric video data fusion by incorporating the architecture and attention model described previously. In addition, we study the ways for integrating different attentions. Finally, we propose an adversarial domain alignment model that incorporates the proposed fusion for unsupervised domain adaptation on multi-modal egocentric video analysis, reaching state-of-the-art performance on the public dataset.

Orthogonal frequency division multiplexing (OFDM) has been regarded as the key technology of broadband wireless communication, and a variety of multi-carrier systems have been developed on this basis. However, almost all multicarrier systems face the problem of high peak-to-average power ratio (PAPR), which brings high bit error rate. Many methods to reduce PAPR have been proposed, such as clipping, SeLective Mapping (SLM), Partial Transmit Sequence (PTS), etc. Among them, hybrid methods can benefit from joint advantages and obtain better performance. This paper proposes an efficient embedded SLM-PTS combination scheme(ESP). This scheme divides the ODFM signal into subblocks, executes SLM steps in each sub-block, and then brings the results of each sub-block to the next step. The simulation results show that compared with the traditional hybrid scheme, this scheme can achieve better PAPR reduction performance and effectively reduce the complexity.

In view of many problems existing in the current railway transportation process, such as pedestrian crossing the guardrail, landslide, rock rolling, debris flow and other factors, this paper summarizes the research status of the underground railway foreign body intrusion monitoring system at home and abroad, and puts forward a railway facility safety monitoring system based on multi-source data from the sky and the sky. The system uses satellite, unmanned aerial vehicle, distributed optical fiber sensor, video, GNSS and GIS technologies to realize the safety monitoring of railway facilities, and formulates the technical scheme for the safety information management of space-based railway facilities, and tests and evaluates the research results through the test section. The development trend of railway facilities safety monitoring system is prospected as follows: foreign body intrusion monitoring at the boundary of key sections; Real-time monitoring of tunnel entrance and high slope section stones; Carry out dynamic monitoring of off-road safety environment.

Since the global eruption of COVID-19, there have been issues with the global medical management system. The classification and treatment of medical waste has brought a huge burden to medical staff. Traditional manual classification has problems such as the heavy workload, difficult identification, and susceptibility to infection. This paper designs an intelligent medical waste detection and classification system based on machine vision. Based on deep learning models such as YOLOv5 and YOLO-resnet18, the data is trained and optimized to obtain a target detection model suitable for medical waste. Through experimental tests, the system has fast detection speed and high accuracy, and the accuracy of medical waste detection and classification reaches 90% and 99% respectively.

The entopallium is a high-level neural structure of the avian visual system that is related to the processing of color- and shape-related information during target recognition. The mechanism underlying it remains unclear and has drawn a considerable amount of research interest. Decoding has invariably been used in relevant studies to verify whether the extracted response-related features contain visual information. However, local field potentials (LFPs) often contain rich features. In this study, We combine the features of different domains to decode color information, including those in the time and frequency domains as well as features related to non-linear and functional network connectivity. To extract valuable response-related features, three types of feature selection methods are compared. The LFP data from the entopallium of a pigeon were recorded during periods of stimulus presentation in which two colors were randomly presented in each trial. The selected feature set was finally used to decode the colors by using several general classifiers. The results showed that the fusion of multi-domain features helped improve decoding performance, and the optimal feature set was a fusion of multiple domains of features. These results improve our understanding of the avian neural mechanisms involved in encoding color-related information.

Coherent Doppler laser velocimetry is an accurate non-intrusive velocimetry method. In this regard, the commonly used methods embrace the fast Fourier transform (FFT) and Weleh spectrum estimations. However, all these methods are limited by the frequency resolution that tightly relies on the sampling points during FFT. This paper introduces a spectrum estimation algorithm using FFT spectral estimation combined with continuous FT spectral thinning, which greatly improves the spectrum resolution and the speed measurement accuracy in the inversion process. Firstly, Hanning window was added to suppress the spectral leakage, and then the rough estimation of Doppler shift was obtained by FFT. Then the spectrum near the Doppler peak obtained by FFT is continuously refined by FT spectrum, and the spectral resolution is increased from 7629Hz to 100Hz. Finally, according to the refined spectrum, the centroid estimation method is used to obtain the Doppler frequency shift, and then the motion velocity is obtained. The velocity accuracy can reach more than 0.0006m /s. The average absolute error of velocity was increased by 51.3%, 65.8%, 77.3% and 83.2% respectively by comparison with Welch spectrum estimation at different speeds of 10m/s, 60m/s, 100m/s and 150m/s. The stability of measurement results increased by 22.2%, 62.6%, 66.9% and 68.2%, respectively. By setting up experiments to obtain the inversion comparison of 10m/s, 30m/s and 60m/s data, the average error of Welch spectral estimation velocity was increased by 39.9%, 24.8% and 67.2%, respectively, and the stability of measurement results was increased by 61.3%, 69.8% and 23.1%, respectively.

Array frequency modulated continuous wave (FMCW) radar is widely used across various platforms to estimate the range and angle of targets using large bandwidth and array angle measurement technology. However, limited platform space and antenna aperture can make obtaining high-precision angle estimations challenging. To address this, a rangeangle estimation algorithm based on spatial and temporal sparse extending is proposed. The algorithm extends the array using a nested array in the spatial domain and a sparse virtual array sampled at different locations through platform movement in the temporal domain. The resulting spatial-temporal extended array is equivalent to a large aperture virtual uniform linear array, improving the angle resolution through difference and spatial smoothing processing. By making full use of spatial and temporal information, this method achieves high-precision angle estimation with fewer array elements.

At present, most unmanned aerial vehicles (UAV) smoke detection systems transmit video back to the ground station computer for analysis to determine whether a fire has occurred, Since the image transmission process takes a certain amount of time and interferes with various interference sources, the response time of smoke detection and the calculation amount of subsequent image processing are increased. In order to reduce the response time of smoke detection, this paper proposes a smoke detection method suitable for UAVs to achieve smoke detection at the UAVs. The improved YUV color model is used to filter and block the video images acquired by the UAVs. Extract the spatiotemporal and dynamic features of smoke; These smoke features are trained and classified using a support vector machine (SVM) to detect the presence of smoke in the video image. Experimental results show that compared with the commonly used smoke detection methods, the accuracy of smoke detection is significantly improved, and the response time is greatly reduced.

As a bridge to connect vision and language, visual relations between objects, visual relation provide a more comprehensive visual content understanding beyond objects. Most previous works adopt the track-to-detect framework for video visual relation detection (VidVRD), which cannot capture long-term spatio- temporal contexts in different stages and also suffers from inefficiency. In this work, we propose a query-based method for video visual relation detection. Our model exploits graph structure to autoregressively generate relation graphs with spatio-temporal contexts and uses an attentional graph convolutional network to fuse the contexts. Experiments on benchmark datasets ImageNet-VidVRD demonstrate the accuracy of our method.

The current image generation research is in full swing, and various image generation technologies emerge in endlessly. Deep learning technology has attracted wide attention since it was proposed. Research on image generation based on deep learning, such as autoencoder network, variational autoencoder network, and generation. Adversarial networks occupy an important position in the field of image generation. However, the generated image effect of each individual image generation network is not ideal. In order to generate clear images, this paper combines VAE (Variational Autoencoder Network) and CGAN (Conditional Generative Adversarial Network) to generate high-quality images. This fusion model has both the advantages of VAE network stability and CGAN network clarity. The experimental results prove that the method in this paper can achieve clear and high-quality images.

Due to environmental changes and other reasons, the types of skin diseases continue to increase, the incidence rate and mortality rate are also rising, and the number of patients is increasing. In addition, people's care for the skin has increased, and the requirements for skin diagnosis and treatment have increased. There are few doctors and many patients, so it is difficult and troublesome to see a doctor. The diagnosis and treatment of skin diseases presents unprecedented challenges. In this paper, the WeChat applet of mobile terminals such as mobile phones is used as the front end to realize the collection, upload and display of information. The server end uses artificial intelligence technology such as convolutional neural network to realize the recognition of skin diseases, build an intelligent diagnosis and treatment system for skin diseases, and realize the diagnosis and scientific treatment suggestions for the patients.

This paper first briefly introduces the open digital cultural museum and the application potential of oblique photography technology in its content construction; then summarizes the principle of oblique photography, and puts forward process flow of oblique photography technology in the specific construction of AR/VR content; finally, it summarizes the shooting skills applicable to different scenarios and shows the modeling results, provide finally Solutions to common problems are proposed so that people can build better 3D panoramas.

Existing image caption generation methods and results mainly focus on identifying existing image contents and their relationships in images, but cannot generate descriptions with fine-grained background knowledge, to overcome the problem that traditional image caption models cannot describe deep-level semantics in pictures. On the one hand, this paper presents an image caption generation method based on target detection and knowledge enhancement. First, in the phase of target detection, this paper presents a Fusion target classification detector (FTCD) which fuses multidimensional information to get tags of human faces, goods, and objects in the graph. Secondly, the knowledge map is introduced, and the related knowledge is queried in the knowledge map using the target tags obtained by the target classification detector. Finally, the target's tag set and related knowledge are jointly fed into the model for coding. On the decoding side of the model, an attention mechanism is introduced to guide the model to select appropriate information and generate an image description. On the other hand, there is a lack of common knowledge about manual descriptions in MSCOCO datasets. This paper presents an evaluation index SPICE-K, which can be used to evaluate image descriptions with common sense knowledge. The experimental results show that the accuracy of the proposed method is 1.3% higher than that of the standard LBPF model. The experimental analysis shows that, compared with the benchmark model, the performance improvement of this method mainly comes from the introduction of the knowledge map and the target classification detector proposed in this paper.

Image segmentation is an important and difficult task in many medical applications. The segmentation results can be used to help doctors diagnose diseases, observe the lesion areas, and make surgical plans. Traditional medical image segmentation methods rely on experienced radiologists, which makes medical image segmentation time-consuming and costly. With the help of computer technology, the use of deep learning neural network for medical image segmentation has been applied in multiple disease lesions or organ segmentation tasks. Cerebral small vessel disease is a manifestation of systemic lupus erythematosus. In this paper, we apply Unet, Nested Unet, and Recurrent Residual Unet to segment the collected medical images of systemic lupus erythematosus patients. The results of Unet and Nested Unet is better than Recurrent Residual Unet. All of the three networks can segment parts of lesion areas, but small areas are missed. The results of our experiments showed the performance of the existing segmentation neural networks on our dataset still need to be further improved to meet the requirements of practical applications.

There are many impurities in the liquid-based exfoliated cell medical pathology slide and the phenomenon of multiple peaks in the evaluation of the front and back images, which may lead to the failure of focus in the automatic scanning and photographing of the system, and the low efficiency of the whole slide scanning. Based on the self-developed digital cell morphology inspection system, a comprehensive evaluation method for the sharpness of fusion image content preprocessing evaluation was proposed, which solved the problem of multi-peak discrimination in image sequence evaluation, and improved the image quality and efficiency of single-point focusing; During the whole scanning, the single-field focusing method of slide based on the traditional mountain-climbing method and the full-field efficient automatic scanning strategy based on multi-point and fixed-plane were designed to greatly improve the scanning speed. The experimental results show that the fusion image content pre-processing comprehensive evaluation method combined with the motion acquisition synchronous background operation strategy has a focus success rate of 98.68%, a single point focus time of 936 ms, and a full slide scan time of 142 s for the 200 field of view, which meets the requirements of clinical liquid-based cell morphology and pathology.

Aiming at the problem of spot center drift caused by the change of medium temperature in the process of laser propagation, this paper considers the influence of actual variable temperature environment, and carries out the research of laser image positioning error compensation method based on sliding linear regression model. First of all, a method to eliminate the gross error of laser image positioning in variable temperature environment is proposed. Secondly, the laser image positioning error method based on sliding linear regression is studied, and the laser image positioning error compensation in the actual variable temperature environment is realized. Finally, the experimental device is built and the field experiment is carried out. The experimental data shows that the peak-to-peak value without error compensation is 3.58 mm, and the peak-to-peak value after compensation is 1.25 mm after comprehensive consideration of the influence of environmental temperature factors, which is reduced by 65%. The experimental results show that the proposed compensation method of laser image positioning error in variable temperature environment based on sliding linear regression model is feasible and has certain engineering application value.

Light weapon systems can play a very important role in fighting under specific conditions (darkness, jungle, canyon, town, etc.). To ensure that light weapons can accurately hit their targets, high-precision optical aiming devices become an indispensable component of light weapons. Generally, light weapons include a precision aiming system consisting of many optical components and precision mechanical parts. Therefore, in the scope high and low temperature and firing stereotypes test, will make the aiming baseline scribe plate and other optical parts become loose, which will certainly make the zero position of the aiming baseline change, the amount of change is the zero position of the aiming device, the amount of change will affect the accuracy of the aiming device, so the measurement of the zero position of the amount of movement is very important to improve the shooting accuracy of the aiming device. For this purpose, we need to investigate a detection device that can be based on digital white micro-optical image processing.

In order to effectively block the spread of the epidemic and meet the needs of epidemic prevention and control in densely populated areas, this paper designed and implemented a real-time mask wearing detection system based on open source data, TensorFlow network model and PyQt5 framework. It has the functions of camera detection and video file detection, which can detect the wearing of masks in real time. It can not only detect the face mask in the image, but also detect the wearing of single or multiple face masks in the video.

In this paper, based on the theoretical foundation of generative adversarial networks and knowledge related to audio enhancement and denoising, a generative adversarial network model and framework are established，and the function of SEGAN audio enhancement algorithm is selected to complete audio noise reduction, improve the purity of audio and realize the audio enhancement algorithm function. At the same time, the training data of the loss function of GAN is optimized from the mathematical principle and the audio enhancement algorithm is evaluated by the speech quality assessment index.This paper replicates SEGAN's audio enhancement generative adversarial network with good practical effects in noise containing frequency denoising and pure audio enhancement, and the enhanced speech of this research method has better auditory quality and intelligibility, and also has better stability compared with the original SEGAN network.

To address the problems of high dimensionality of hyperspectral images, small training samples and overfitting and too many parameters caused by model training, a hyperspectral image classification model (CBAM-HybridSN) with improved hybrid neural network combined with convolutional attention mechanism is proposed. The model firstly uses principal component analysis to remove the redundancy of spectral dimensional data, extracts the null spectral features by the hybrid neural network model, and introduces the convolutional attention module to rescale the extracted features and highlight the important features, thus improving the classification accuracy. In the experiments, the Pavia University dataset was divided into samples with 1:9, and the OA reached 99.32%, achieving accurate classification of hyperspectral images with small samples.

As eggs are a very important food source in our daily diet, the inspection of egg quality must be strict. In this paper, a detection method based on machine vision image processing is proposed for the detection of eggshell cracks. Through image pre-processing, the images of the three BGR channels are compared, and then the median filtering is used to derive the image information of the R channel that is more favorable for detection. The image is enhanced to find the ROI, and connected region and to perform image segmentation to get the crack information. Finally, a morphological operation is used to extract the feature value of crack length from the image and conclude that the crack in the eggshell is detected by judging the feature value of the image. The experiments show that the accuracy of the OpenCV image processing-based detection method can reach more than 90%, which indicates that the method of detecting eggshells through image processing has unlimited potential.

Aiming at the problems of low detection efficiency and poor detection accuracy of traditional steel surface defect detection methods, this paper proposes a steel surface defect detection algorithm based on improved YOLOv5 network. First, the lightweight GhostNet network is introduced, and the Ghost module is used to optimize the YOLOv5 backbone feature extraction network to obtain a lightweight model to reduce the complexity of the model and improve the speed of crack detection; Then, the efficient CA (Coordinate Attention) channel attention mechanism is fused at the model prediction output to further enhance the defect feature extraction ability and improve the model detection performance. The simulation results show that the model volume of this method is reduced by 10.3% and the mAP value is increased by 4.7% compared with the existing YOLOv5 algorithm. Compared with traditional steel surface defect detection methods, the proposed algorithm can detect the types and locations of steel surface defects more accurately and quickly, and has a smaller model volume, which is convenient for deployment in mobile terminals.

Imaging of static scenes under water is widely used in underwater detection technology. Aiming at the problem of severe distortion of underwater scene images caused by water surface fluctuations. An underwater distorted image restoration algorithm based on compressed sensing and centroid method is proposed. First, track the significant feature points between video frames to obtain the displacement trajectory of the feature points, then use the compressed sensing algorithm to reconstruct the motion vector field of the entire image to eliminate the periodic distortion of the image. Further, on this basis, the dense optical flow between the images is calculated to construct the centroid of the image sequence, thereby eliminating the randomness distortion that may exist in the image. The experimental results show that compared with the current mainstream restoration methods, the algorithm can ensure that the restored image has high definition while eliminating image distortion, and it performs well in some image quality evaluation indicators.

Building change detection from high-resolution images plays an essential part in landcover classification and the analysis of urban spatial information. With the development of deep learning (DL) in remote sensing, DL has become the mainstream method in the building change detection field. However, the current building change detection methods still only can detect building change, but few studies investigate whether the building is increasing or decreasing, in other words, they can only capture the spatial change of buildings, but can’t get the temporal change of buildings. To solve this problem, this manuscript proposes a new simple but effective multi-task learning building change detection method, which can capture building increase or decrease on high-resolution images at the same time. First, we modify the DenseNet121 as our backbone to extract deep semantic features, mainly because DenseNet has a strong feature extraction ability, but the computation cost is relatively low. Second, we design a differential structure to detect building increases or decreases at the same time with a new loss function, namely Tanimoto loss. A large image with very high resolution is selected to verify the effectiveness and robustness of our method. On the one hand, the result indicates that our proposed building change detection method can detect building increases or decreases accurately, and the quantitative precision shows that the F1-score is over 80%. On the other hand, this paper presents model design idea for building change detection model, which can provide several novel perspectives to other research fields, such as object detection on high-resolution images, change detection of other objects.

Defect panoramic ultrasound image is helpful for material integrity evaluation in NDT&E. However, the single scan area of ultrasonic transducer is limited, which leading to the limitation of image area. Due to the low resolution and minimal detail in ultrasound images of defect, traditional visual image stitching methods cannot be directly applied to ultrasound defect images. Therefore, an ultrasonic image stitching method for material defect detection is proposed. According to the feature analysis of ultrasound defect images, an improved scale-invariant feature transform algorithm for ultrasound images stitching is developed. The algorithm first identifies image feature points in the same horizontal layer for coarse matching and then uses Euclidean distance histograms to eliminate mismatched pairs for fine matching. Thereafter, the algorithm transfers different images into the same coordinate system by using a transformation matrix obtained from a random sample consensus algorithm with an adaptive threshold. Finally, an improved fade -in/fade-out weighted fusion algorithm is used to stich the images. Acoustic field simulations and actual test experiments have been performed on an aluminum test sample to obtain a panoramic ultrasound defect image by using the proposed method.

At present, due to the complex texture, small defect and large scale span of some surface defects in aluminum profile surface defect identification, the detection accuracy is low and fails to meet the requirements of industrial detection. To solve these problems, an improved Yolov5 algorithm model is proposed. Based on the Yolov5 model, the BOA3 (Bottleneck PfAAM) module is used to replace the previous C3 module for small targets, which tend to improve the ability of high-level feature extraction, and reduce the number of references and the amount of computation. For multi-scale problems, the middle C3 module of the main trunk is Bottleneck changed to a double-branch structure, which tends to fuse characteristic information of different scales. Finally, the loss function CIOU is replaced by EIOU, and EIou-NMS is introduced to remove redundant boundary boxes and improve the inference accuracy of the model. Experimental results show that the detection accuracy of the improved Yolov5s model is 9.8% higher than that of the original model, which is superior to the latest Yolov7 model.

Significant progress has been made with image super resolution using deep learning in recent years. However, existing methods aim at minimizing differences between pixels, which means they fail to consider the distribution difference while reconstructing images. To address this challenge, we propose DWGAN: a dual Wasserstein-Autoencoder based generative adversarial network. Our DWGAN successfully implemented the adversarial training process both in the latent space and the image space. In the latent space, the adversarial training process successfully models the latent distribution with Wasserstein Autoencoders and the adversarial training in the image space achieves better reconstruction performance. Besides, we also utilize tree-structure residual encoders to extract 2-dimensional latent features of the same size as low-resolution (LR) inputs, which enables pixel-wise matching. Experiments show that our DWGAN achieves better visual and quantitive results compared with other GAN-based models.

Hyperspectral remote sensing can synchronously obtain the surface coverage space image and continuous spectral data, and can realize fine classification and recognition of ground objects. The motivation for this paper is to search an effective specteal feature extraction method for hyperspectral classification. The main methods used to extract spectral features of hyperspectral images are principal component analysis (PCA) and linear discriminant analysis (LDA). The realization of classification based on k-nearest neighbor (KNN) classifier is discussed. The training set and test set of Pavia University, Indian Pines and Salinas are selected as the data source, and the selection experiment of algorithm parameters is set, and the classification performance under different feature extraction methods is compared and analyzed. The experimental results show that with the parameters of OA, PA and Kappa based on the confusion matrix as the evaluation indicators, the classification accuracy based on LDA method is higher, and the effective classification of ground objects can be achieved. The results could provide a basis for further research on hyperspectral image classification based on spatial spectral feature extraction.

Bloodstains, as a biological test material with a high occurrence rate at the scene of a criminal case, can provide a great deal of clue to the case. How to quickly, accurately and non-destructively distinguish between bloodstains and suspected bloodstains in a crime scene and identify the species of bloodstains is not only necessary for the examination of physical evidence in judicial practice, but also an inevitable requirement for the public security authorities to identify the case and reveal the falsified scene. Using a hyperspectral imager, we acquired hyperspectral data from 26 experimental materials, including 13 animal bloodstains and 13 suspected bloodstain compounds. This paper focuses on the identification of bloodstain species. The experimental technique builds a classification model for the identification of bloodstain species using three algorithms: K-nearest neighbor, Support Vector Machine, and Naive Bayes classification. The experimental results showed that the Naive Bayes classification algorithm and the Support Vector Machine classification model with RBF kernel function and Sigmoid as kernel function were the most efficient. Experiment classification accuracyof100.00% was achieved in all cases. Our experiments explore a new method for identifying different species of bloodstains using hyperspectral imaging technology.

This paper addressed the contradiction between the decrease of labor force and the increase of agricultural output demand in agricultural development, combined with the technical advantages of UAV imaging radar in SAR imaging, ground moving target detection and height measurement, analyzed the key technology of UAV imaging radar applied to agricultural plant protection. The signal processing scheme and processing process of UAV imaging radar for agricultural plant protection were designed, and the experimental verification results using the measured data were reasonable and feasible. The verification results of the measured data showed that the UAV remote sensing imaging technology can realize the high-resolution imaging and preliminary ground object classification in the operation area.

Spectral images include rich spatio-spectral information of target scene, which can accurately identify and distinguish the features of ground objects. Therefore, spectral image classification is widely used in remote sensing. However, traditional spectral imaging techniques need to scan the region of interest along the spatial dimension or spectral dimension, which takes a long acquisition time and increases the burden of data transmission and storage. To overcome these shortcomings, coded aperture snapshot spectral imaging (CASSI) system based on compressive sensing theory appeared. In this paper, we build a testbed of dual-disperser CASSI (DD-CASSI) system, which can reconstruct the three-dimensional (3D) spectral image datacube of target object from a few two-dimensional compressive measurements. Then, a 3D convolutional neural network is applied to accomplish the spectral image classification based on the reconstructed datacube. Different classification methods are compared based on the experimental data. It shows that the proposed compressive spectral image classification method achieves pretty close results compared to the classification methods based on the original datacube. But, the proposed method is beneficial to improve the acquisition efficiency of the spectral image data.

Face spoofing detection technology can accurately distinguish the truth of the face captured by the camera, and it has been applied in many security fields. In view of the problems of the existing face spoofing detection models, such as the single way to extract image features and the insufficient semantic information, a model architecture that fused the predictive feature maps of different scales was proposed. Firstly, the face image is randomly cropped and the local image is taken as input; Then, a cheap feature predictive convolution attention module is designed to alleviate the problem of feature map redundancy, and make full use of the similarity features of adjacent blocks around local pixels, that is, predict the feature value of the central region through surrounding pixels; Finally, the feature maps extracted from different levels are fused. The model aims to achieve high accuracy in identifying the authenticity of face images. The experiments show that the accuracy rate on CASIA-SURF (Depth modal) dataset is 99.42%, the average classification error rate is 0.53%, and the zero error rate is achieved on CASIA-FASD and Replay-Attack datasets. The model parameter quantity is only 0.37M, which is lower than most of the model architecture based on convolution neural networks.

The existing interior design platform can draw interior design three-dimensional images according to the shape of interior objects. On the basis of existing research, according to the location of indoor objects, generate real and threedimensional images, and design an indoor design platform based on three-dimensional virtual vision. Determine the structure of the interior design platform by analyzing the three-dimensional perspective, and complete the interior positioning algorithm and interior design demand analysis; By improving the interactive functions of the interior design platform, the anchor points and drivers of the design platform, the interior design platform based on 3D virtual vision is built. The simulation platform uses the environment design contrast experiment, the result shows that the interior design platform based on 3D virtual vision can draw the interior design plane image and three-dimensional image, and the position of the interior objects in the image is more in line with the actual needs.

Target radiation noise preprocessing is the basis and premise of underwater acoustic signal processing. In order to obtain better noise reduction effect, a method of underwater target radiation noise reduction based on superposition variational mode decomposition residual is proposed. The initial signal collected by passive sonar is decomposed into a series of intrinsic mode functions and residuals by variational mode decomposition, and the spatial-dependence recurrence sample entropy of each intrinsic mode function is calculated. According to this value, the intrinsic mode functions are divided into noise-dominated intrinsic mode function and signal-dominated intrinsic mode function. The noise-dominated intrinsic mode function is denoised by improved wavelet threshold denoising. While SG filter denoises the signaldominated intrinsic mode function, restructures the processed natural mode function, and superposes the residual from the variational mode decomposition to obtain the de-noised signal. The proposed method was applied to the experimental data of SWellEx-96, an open source data set. The results show that, the proposed method has certain advantages. The improved signal-to-noise ratio is higher than 55.4dB and the root-mean square error is less than 1.8*10-5, which makes the line spectrum characteristics in the frequency domain more obvious. Therefore, this method has certain practical application value.

The current conventional airport road surface disease detection method mainly obtains disease information data and analyzes them, which leads to poor detection accuracy due to the lack of denoising of the disease images. In this regard, the visualization detection method of airport road surface disease based on BIM and CCD imaging is proposed. Combining the BIM forward model and CCD imaging technology to collect the disease information, the image is denoised and filtered, and the edge detection algorithm is optimized. The experimental results show that the proposed detection method has a high AP value and a more desirable detection accuracy.

In recent years, the field of thermal to visible image transformation has made great progress with Generative Adversarial Network(GAN). Since GANs own the ability to fit the signal frequency, these methods are limited to using more GANs or designing sub-tasks to improve the quality of the generated image. Due to the large difference in the information frequency of individual objects in the image, GANs pay less attention to low-frequency information, which leads to edge distortion of dynamic objects such as people and vehicles in the generated image. Edge distortion has become an important issue for practical applications We propose a new structure and two new loss functions to complete the task of thermal to visible image transformation. Our network takes one thermal image as input and generates a visible RGB image as output. First, we down-sample inputs for three times to get multi-scale inputs. Then we build outputs in multi-scale form small size to large size and fuse the multi-scale outputs. In train stage, our two new losses better direct the network to converge. In order to prove the effectiveness of the proposed thermal to visible image transformation algorithm, we use the KASIT-MPD dataset, which includes real paired thermal and visible images.

Noise source localization is very important in the field of acoustic measurement. However, the localization accuracy in strong reflection conditions using traditional localization method will be seriously disturbed by the sidelobes. In this study, a new noise source localization method of multi-frequency energy superposition (MES) was proposed to reduce the effect of sidelobes. The noise source maps in strong reflection conditions obtained by MES method were analyzed and compared with delay and sum method. The results show that the main-to-sidelobe suppression ratio of source localization obtained from MES method is higher than that from traditional delay and sum method, indicating that the proposed MES method is more accurate for noise source localization in strong reflection conditions.

In recent years, deep learning-based methods have made a big splash in the field of single image super-resolution (SISR), and a lot of excellent research works on lightweight image super-resolution networks have emerged. This series of approaches point a way forward for lightweight SR network design, but they generally suffer from two problems: 1) slow convergence of the network 2) poor image reconstruction quality. In this paper, we firstly propose a simple and effective learning rate decay scheme without additional cost, and the new strategy can accelerate the convergence speed of the network, and the model obtained in the same training time is about 0.05 dB higher in PSNR test results. Secondly, we think about the performance bottleneck of RFDN and propose an enhanced full-channel attention (EFA) module with a mixture of channel and spatial attention, which can improve about 0.085 dB in PSNR compared to the CCA attention module. Finally, we propose a new feature distillation network based on the EFA module, called EFAFDN. With only 523k parameters, the EFAFDN(x2) is lighter than the RFDN with 534k parameters and has better SR performance than the RFDN-L with 626k parameters. Extensive experimental results show that EFAFDN performs slightly better than RLFN, the winning model of the main track of the NTIRE 2022 Efficient Super-Resolution Challenge.

With the increasing use of urban vehicles, intelligent transportation has also attracted more and more attention. License plate location technology is a very important technology in intelligent transportation. First, the interference factors and noise in the image are removed through image preprocessing, and then the license plate location is realized through rough location and accurate location. Through rough location, the license plate position can be initially positioned. Through accurate location, the license plate area can be corrected to achieve accurate location. The domestic and other regional license plates are selected for experiments. The experiments show that the method can effectively locate the license plate area accurately.

Stereo matching is a critical step in three-dimensional reconstruction. The conventional Census transform is excessively dependent on the central pixel and is susceptible to noise interference. Therefore, this study proposes a cost calculation method for multi-feature fusion based on the HSV color space and an improved Census transform. To address the problem of poor matching in regions with weak textures and depth discontinuities. In this study, a new window construction method is developed based on the length and variable color threshold to obtain the crossover region. After dividing the weights of the constructed windows, the cost aggregation is calculated. Finally, a dense parallax map is obtained via parallax calculation and optimization. Experiments were conducted based on the data provided by the Middlebury dataset. The experimental results showed that the improved cost calculation algorithm was significantly resistant to noise, and the average mis match rate of the algorithm was reduced to 8.03%.

In view of the text in image or video on the visual interference and privacy issues, a text removal method combining text location and image inpainting is proposed. Firstly, the method inputs the image with text into the text location network to identify the text contour area, and then generates the masked image to simulate image damage part according to the text contour area. Finally, the damaged image is repaired through the damage reconstruction network to realize text removal. At the same time, in order to ensure the reliability of the text removal, the complementary fusion layer is proposed to ensure the non-text area in image do not change. The experimental results show that compared with the traditional texture and patch repair algorithm our method has a better repair effect. Compared with the improved generative adversarial network method, our method is more close to the original image features.

DDGI, GIBS, and other similar dynamic global illumination algorithms solve the irradiance integral by shortening and reusing the light path, obtaining the multi-bounce diffuse reflection effect without recursion. However, these algorithms may perform poorly in scenes with complex lighting. GIBS introduced a simplified path guiding algorithm, which uses lighting information from history, and is limited by the resolution of radiance maps, causing inaccurate sample pdf under dynamic lighting. We propose an importance resampling method based on surfels that is more adaptable to scenes with dynamic lighting and provides more converged diffuse integral results. First, we derive the radiance composition under next event estimation(NEE) framework, then gather visibility maps and calculate the direct irradiance additionally to obtain the target pdf. Finally, we build reservoirs for each ray used for diffuse integral, and generate ray samples by reservoir resampling. We also provide a feasible temporal reuse strategy for single and persistent reservoirs based on surfels. Experimental results demonstrate that our method can generate higher quality samples, improving the accuracy and stability of diffuse global illumination under dynamic lighting.

Understanding protein subcellular location in immunohistochemistry (IHC) images is conducive to investigating textured background at the protein tissue level and analyzing the interaction mechanism of the protein molecule. At present, the mainstream IHC image classification models are mainly designed based on classical machine learning and general deep learning, which expose the shortcomings of gradient dispersion, low information richness, and weak discrimination. In this paper, the IPA net is developed to classify IHC images based on Interactive Pointwise Attention (IPA), whose advance can be concluded in the following: firstly, patch splitting is adopted to split IHC images to dilute tissue diversity, bubbles, noise and other interference factors; secondly, the IPA unit consisting of interactive computation is proposed to modify pointwise pixels and enrich the feature richness. The experimental results demonstrated that IPA net reaches97.33%, 97.28%, and 97.24% in accuracy, precision, and recall, respectively, which out perform other baseline Convolutional Neural Networks (CNNs). Moreover, feature maps were visualized to investigate spatial distribution information, which provides a theoretical basis for analyzing protein subcellular pattern.

Accurate segmentation of the hippocampus from MR brain images is a critical step for investigating early brain development. Unfortunately, the previous methods hippocampus segmentation is not suitable for subdividing hippocampus into multiple anatomically different subregions, such as head, body, tail, based on structural magnetic resonance imaging. The main problem is lack of discriminative and robust features representation for distinguishing the hippocampus from the surrounding subregions structures. To this end, we designed a dual-pathway network to learn the global and local representation from both global and local pathway. We proposed a global pyramid attention fusion module (GPAFM) in global pathway, which extracted coarse and global features representing the gist of the object. And a local feature aggregation module was adopted in local pathway to extract fine and local features representing the details for determining object categories. The experimental results have demonstrated that GPAFM successfully parcellated the hippocampus into head, body and tail subregions with high accuracy. Dice coefficient on three hippocampus datasets have further demonstrated the validity of the model. Compared with superior methods, the proposed method had better performance on parcellating the hippocampus subregions.

Automatic segmentation of targets from medical images is a difficult task because of the high complexity of medical images and the lack of simple linear features. U-Net is widely used in the field of medical image segmentation for its state-of-the-art performance. However, there are still many problems, the most serious of which are the lack of accuracy of image segmentation and the large amount of information loss. To alleviate these deficiencies, we propose a UNetbased image segmentation model GAM-UNet, which has the same coder-decoder structure. The Global Attention Mechanism (GAM) is introduced and embedded into the encoding part of the proposed model to automatically adjust the weights of the feature maps. Evaluation results on the public datasets Data Science Bowl and CVC-ClinicDB show that our model performs better than other U-Net variants. After additional analysis, our proposed model is more hardware friendly and allows for faster medical diagnosis with a slight increase in accuracy. Our proposed approach can also be extended to other image processing tasks, such as object detection and tracking.

Stress refers to a series of physiological reactions to help the human body overcome threats. In this study, the machine learning method was used to analyze the stress in students' daily life environment. The purpose is to explore the neurophysiological model of sleep under stress, so that they can effectively adjust their personal stress and improve their sleep quality. First, we collected ECG and triple-axis acceleration data of 33 college students in their daily life, calculated the inter-beat intervals from every two consecutive R waves in these data. Secondly, we extracted 39 linear and non-linear RR parameters as stress physiological features, and applied Leave-One-Subject-Out cross test and Sequential Backward Selection algorithm for feature selection while training some traditional classifiers. Finally, a binary classification model of stressed and non-stressed states recognition in the first hour of sleep at night was constructed, and the generalization accuracy of the model was 75.00% on the validation data set independent of model training and feature selection. The results show that it is feasible to monitor students' stress state at night by machine learning.

The Memristor has become one of the best devices for hardware realization of neural networks due to its nanometer size, low power consumption, fast reading and writing speed, and variable resistance. This paper proposes a memristor crossbar structure for processing 3D convolutional neural networks, which can realize the function of action recognition. The 3D convolutional neural network can extract features from the space and temporal dimensions simultaneously, thereby obtaining the coding information in multiple adjacent frames, and is widely used in video processing. At the same time, based on the advantages of the 3×3×3 convolution kernel, we built a memristor crossbar to realize the function of video recognition. This structure has good flexibility and anti-interference ability and can reduce floor space. Finally, we implemented the network on the Weizmann datasets and achieved 85% accuracy. The results show that the Memristor can realize the task of classifying human action behaviours and can realize more extensive and complex neural networks to process complex applications.

The majority of non-contact heart rate detection are based on pure traditional or pure deep learning methods, which struggle with high noise interference and low accuracy in complex scenes or with insufficient training data. Therefore, this paper proposes a hybrid method that combines traditional signal processing with deep learning, leveraging the complementary advantages of both, to enhance heart rate detection accuracy and robustness. First, Gaussian heatmaps are used to predict the facial keypoints for Region of Interest (ROI) localization, from which the skin color change signals of RGB channels are extracted. After using Fast-ICA to demix the pixel averages of the three channels, three independent source signals are obtained, and the signal with the highest signal-to-noise ratio is selected as the remote photoplethysmography (rPPG) signal. Then Gramian angular field method (GAF) is used to encode the feature image of rPPG signal as the input of LA-Res2Net, and finally the model can infer the heart rate after training. The effectiveness of the proposed method is verified using the VIPL-HR-V2 dataset. Experimental results show that compared with traditional image processing methods and other advanced models, the proposed method reduces the mean absolute error (MAE) from 15.2 BPM to 6.2 BPM, the root mean square error (RMSE) from 19.5 BPM to 9.1 BPM, and increases the Pearson correlation coefficient (R) from 0.22 to 0.62. This indicates that the proposed method provides a reliable solution for remote non-contact human heart rate detection.

In the new media era, the innovative application of a large number of digital information technologies has promoted the upgrading iteration of the media, and also prompted people's visual aesthetic needs to change. In order to solve the problem that the traditional visual communication design is flat and passive, this paper takes the visual communication effect as the research object, takes the virtual reality visual interactive experience technology as the core, and combines the network information technology and computer application technology to build a three-dimensional visual platform for visual communication design, which improves the artistic expression effect of visual communication. The platform will use 3ds Max, Photoshop, Unity3D and other software to transform the three-dimensional model of traditional visual communication design works, and add a large number of visual interactive operations, and then integrate and package them by Javaweb technology to form a web application program, so as to fully realize the integration of virtual reality technology and visual communication design. The measured results show that the three-dimensional visualization platform is of positive significance to the improvement of visual communication form and promotes the innovative development of visual communication design.

Research on gas sensor drift compensation for electronic noses can be traced back as far as the 1990s, but to this day, the gas sensor drift problem is still a thorny issue. Research on sensor drift compensation algorithms is still indispensable. Most of the traditional methods of electronic nose drift compensation operate from the aspect of signal processing, but this method has limited application environment, while nowadays more compensation is performed from the perspective of pattern recognition. Convolutional neural network, as a classical deep learning model, can extract not only samplespecific features but also domain-invariant features between different domains. In this paper, based on the above theory, we propose a novel drift compensation model based on convolutional neural network by combining the characteristics of real-world mammalian olfactory system, which is designed with multi-branch multi-classifier convolutional network structure to do specificity feature extraction on gas data and make the network model have better generalization and flexibility. Experiments on a short-term drift dataset show that the multi-branch convolutional neural network proposed in this paper has better performance.

Current studies suggest that one-shot multi-object tracking approaches are as competitive as methods that separate inference into two branches. However, current methods neglect the inherent differences and relationships between tasks. To address this question, we propose RoTrack, which incorporates a novel inference network and efficient association approach. Our network module learns both inter-task associations and difference to alleviate inference inequality. We introduce the coordinate associating opportunity (CAO) strategy, which employs fair voting to avoid errors resulting from manual selection of association bases. We evaluate our algorithm on public MOT Challenge datasets and demonstrate outstanding results. Experiments show that RoTrack can well deal with the above-mentioned problems in the field of multi-object tracking.

With the development of intelligent monitoring and autonomous driving technology, the adverse effects of rainy weather are receiving increasing attention. The current image de-raining techniques have two main drawbacks: first, they easily lose local features and channel features, and second, the computational cost of global calculations on large feature maps is staggering. In this paper, we propose a self-attention network called URformer, which integrates spatial features and channel features. We introduce a novel feature extraction module that effectively combines local self-attention and channel self-attention mechanisms. URformer significantly reduces the computational cost of traditional global Transformers while extracting and preserving channel and spatial features. Experimental results demonstrate that the URformer network exhibits excellent performance on the RGB image de-raining task.

In order to solve the problems of color distortion and blurring common in underwater images, this paper uses color channel compensation and grayscale equalization to correct color cast, and uses the total variation of the reweighted graph to estimate the fuzzy kernel to achieve image restoration. Based on the average gray value of the color channel, the color channel with serious information loss is compensated to reduce the deviation of the gray distribution between the color channels, and the gray distribution balance is further realized by segmented stretching. Combined with the reweighted graph total variation prior algorithm and the improved L-R unblinded restoration algorithm to restore the final image to remove blurring. Several sets of different data sets were selected to compare the proposed algorithm with several classical algorithms, and the evaluation results of subjective vision and three objective indicators were analyzed. The results show that the proposed algorithm can be applied to underwater image restoration, and the algorithm has good general applicability, which can effectively correct color cast, remove blur and improve image clarity.

The key point of this thesis is to solve different resolution facial expression recognition. Previously, we used the Highresolution image generation module (HIG) and Coarse to fine the classification module to solve this problem. The effect of HIG is to convert low-resolution images to high-resolution images through superresolution. Coarse to fine classification module instead analyze the image step by step using a three-paragraph classification. Both methods can work well.

In recent years, the imaging theory and technology of missile-borne synthetic aperture radar (SAR) has become one of the important research topics. The missile-borne platform usually requires high-precision imaging from the front and side view, and the missile is affected by the airflow, and the flight trajectory changes greatly and randomly during the air flight. For this reason, a motion compensation algorithm is required. At the same time, due to the real-time and large computational requirements of missile-borne SAR imaging, it is of great significance to use a higher-performance graphics processing unit (GPU) platform for SAR time-domain imaging processing. This paper introduces the use of GPU to implement extended Wavenumber-domain (EOK) algorithm and large-squint motion compensation algorithms.

At present, autonomous vehicles equipped with night vision systems have been widely used, which can provide drivers with visual aids and collision warning at night or in low light conditions, so as to reduce the probability of accidents. In this paper, the image quality evaluation indicators and test methods of automotive infrared night vision camera have been studied. According to the characteristics of infrared night vision camera, NETD, MRTD, MTF and identification distance have been proposed as the evaluation indicators, the test methods and procedures have been given. A prototype was selected for verification test. The experimental results were accurate, objective and reproducible, which could be used as the test and evaluation method for automotive infrared night vision camera.

Aiming at the problem of template degradation and low real-time tracking in the scene of target jitter and target pose change in the process of UAV extended target tracking, an edge template matching method is proposed in this paper. Specifically, the edge information of each frame image is utilized to update the target edge template by integral and enhance the robustness of the template. Besides, FPGA is adopted to realize the hardware circuit design of the proposed algorithm and optimize the tracking real-time problem. Experimental results demonstrates that the proposed method have satisfactory performance, the tracking matching rate reaches 90% and the tracking speed reaches 200 FPS.

Due to the strong nonlinear characteristics of vehicle system, obtaining the current vehicle motion state accurately is the basis for improving vehicle control accuracy. Moreover, with the increase or decrease of passengers and goods, some parameters of the vehicle will change, and accurate values cannot be obtained at any time. In this paper, the Dual Unscenter Kalman filtering algorithm (DUKF) is compared with the Unscented Kalman filtering algorithm (UKF) by using the vehicle dynamics grey box model in Matlab. The results show that DUKF still has a certain estimation accuracy when estimating vehicle state and parameters.

Improving the survival probability of important ground facilities requires setting protective measures above ground with the development of precision guided weapons. For low altitude protection, a suitable detection system for a given velocity of target is a must. A special detection system fulfilling the requirements for protection was constructed based on multiUAVs coupled with radar device to track the coming target in this work. The target’s motion model was built as a function of signal arrival time, which was obtained from the objective function. To achieve better stability and tracking performance., an optimization method was presented according to the arrangement of detection system and parameters of the target, by minimizing the bias between the system centre and target position. It was found that effective tracking was achieved for the non-cooperative target with following control taken into consideration.

The smoke and particulate matter emitted from forest fires will have a significant impact on air quality and climate change, as well as human health and the ecological environment. Fire smoke can be detected from satellite imagery based on changes in its spectral properties. However, previous remote sensing satellites used for smoke detection have low to medium resolution and long revisit intervals. In response to this deficiency, this paper proposes a method for extracting fire smoke based on the multi-channel threshold of GOCI-II (Geostationary Ocean Color Imager-II) data through the analysis of the fire smoke plume sensitive band. Firstly, the separability index of reflection of smoke and other objects in the ultraviolet, visible, and infrared bands was analyzed. Experiments have shown that the Rayleigh scattering of smoke is stronger in the ultraviolet and blue bands. This difference can be used to separate smoke from clouds. Finally, we compared the monitoring results of the multi-band threshold method with the SSDA monitoring results of VIIRS. Through comparison, it is found that the multi-band threshold method is better for smoke plumes extraction and can obtain more detailed and comprehensive smoke information.

The detection technology of infrared thermal imaging is becoming more and more mature. With the wide application of deep learning, the accurate detection method of infrared image power components based on object detection occupies the mainstream of the market. The existing accurate detection methods of power components mainly have problems such as difficulty in taking into account the speed and accuracy, and low recall rate of rotating targets. This paper proposes a Lightweight Dual Scale Angle Transform Network (LDS-ATN). Based on the lightweight backbone network, an angle transform module is introduced to enhance the angle correction. Finally, this paper achieves an average accuracy of 89.89% and a detection speed of 127.27 FPS on the task of accurate detection of electrical components. It proves the practicability and generalizability of this network in power detection.

To solve the tracking drift caused by rotation, uneven illumination, and beyond the field of view in the moving process of the target with low contrast extension, a method of edge feature points matching is proposed in this paper. Specifically, feature points are extracted from image edges to improve the stability of feature points matching. In addition, combined Non-Local algorithm and improved Contrast Limited Adaptive Histogram Equalization algorithms are used to enhance the image. Experimental results show that this method has satisfactory performance, has good anti-rotation characteristics, can be stably tracked when the target reappears in the field of view and basically meets the real-time performance.

Leukemia is an abnormal proliferation of white blood cells in bone marrow and blood. The abnormal increase of immature white blood cell count and the decrease of other blood cell counts may be signs of leukemia. Usually, pathologists diagnose leukemia by observing blood smears under a microscope. Bone marrow examination can be recommended to confirm and identify the specific type of leukemia. These routine methods are time-consuming and may be affected by the skills and expertise of the medical staff involved in the diagnostic process. Image processing based methods can be used to analyze microscopic smear images to automatically and quickly detect the incidence rate of leukemia. Image segmentation is one of the important tasks in processing and analyzing medical images. In this paper, we try to review the relevant works in the medical image processing field of blood smear images, focus on the automatic detection of leukemia, and classify and discuss the works in the relevant fields according to the segmentation methods used.

The mid-deep geothermal resource is one of the important clean energy sources for winter heating in northern cities. Geophysical methods play an important role in geothermal exploration, but the effect of geophysical methods is not good because of various types of serious interference in urban areas. This paper introduces a new high-power time-frequency electromagnetic technology and its geothermal exploration test in Dongli District of Tianjin. The test proves that this method has the characteristics of strong anti-interference ability, high signal-to-noise ratio and high construction efficiency. It can collect objective and real data in geothermal exploration in urban areas with stratigraphic physical properties. Through fine data processing, the stratigraphic structure and karst development in the test area have been ascertained, the location, trend and nature of Cangdong fault and its secondary faults have been ascertained, and five favorable geothermal areas have been delineated, which provides geological basis for deep geothermal resource evaluation and drilling.

Osteoporosis is known as the "silent killer" and some of its complications have a higher mortality rate than cancer. The current method of diagnosis, X-ray absorptiometry can cause radiation damage to the body. Ultrasound guided wave is sensitive to a wide range of bone information, and is harmless to humans, which is a new solution to evaluate bone condition. In the cross study of medicine and engineering, there are always compatibility problems when engineering equipment is applied to medical research objects. Under the background of osteoporosis detection, this paper explored the application of ultrasonic guided wave detection technology in the measurement of long bone thickness. In the experiment, a sample support platform was designed to solve the problem of fitting the guided wave transducer with the bone surface. The platform consisted of an upper plate, springs, screws and a lower plate. The sample can be stably supported by the adjusting screws, and the guided wave transducer can be adapted to the irregular sample surface by the pre-deformation and elasticity of the springs. The thickness of cortical bone was successfully measured in this experiment. The error of tabular cortical bone was 0.91%, and that of tubular cortical bone was 0.41%, which showed that the experimental method and data processing method were reasonable and effective. This method can provide basis for clinical diagnosis of osteoporosis in the future.

Water surface object detection is one of the essential techniques for carrying out water-related tasks and assisting in the conduct of vessels. However, water surface object detection algorithms currently have low-precision problems detecting small and irregular objects, and difficult samples are hard to detect in harsh environments. To this end, an improved loss and enhancement feature fusion method is proposed to optimize the YOLOX object detection algorithm for water surface target detection. YOLOX, as an anchorless frame algorithm, has better detection capability for irregular objects. Improved loss improves the focus on difficult samples and treats positive samples asymmetrically based on EIoU loss. The Lite-RFP enhanced feature fusion mechanism enables the network to recursively pass contextual information recursively. It enables the shallow network to integrate deep semantic information better, improving the performance of small target detection and maintaining a lighter network structure. The experimental results show that the improved algorithm based on YOLOXs improves the mAP value by 4.11 percentage points compared to the original YOLOX-l. At the same time, a real-time detection speed of 56.534 FPS can be achieved, and the detection problems caused by different water surface environments are improved.

In practical applications, remote sensing (RS) scene classification faces data shift problems, including novel class and data discrepancy problems. Due to these problems, it is difficult to obtain representative and discriminative features. Therefore, we propose improved prototypical networks (IPN) based on few-shot learning to solve data shift problems in RS scene classification. First, a novel and effective scheme is proposed, which is to introduce Vision Transformer (ViT) pre-trained on a large-scale dataset as the feature extractor of the prototypical networks. Based on the meta-task training framework, IPN can adapt well to RS scene classification tasks and obtain representative features. In addition, a novel loss function called self-distillation-based prototype loss is designed to obtain discriminative features by introducing inter-sample selfdistillation and inter-layer self-distillation methods. Extensive experiments are conducted on several public RS scene datasets. Compared with the existing methods, the proposed method achieves an improvement of 4.18%-21.42%. Results demonstrate that the proposed method can effectively solve the data shift problems in RS scene classification.

Adversarial examples have raised public concern about the robustness of deep neural networks (DNNs). One universal approach to enhance the robustness is adversarial training which essentially augments the training data. However, adversarial training succeeded only to a very limited extent. This limited progress is partly due to the lack of interpretation and understanding of the robustness of DNNs. In this context, we try to explain the adversarial robustness by embedding the sample points onto a hypersphere, which naturally provides an interpretable metric for the distance between sample points. Different from the empirical intuition, it is observed that adversarial trained models show complex patterns when facing different datasets and training configurations. Observations and explanations about robustness and model behavior are made from the aspect of the distances between samples points. Lastly, we discuss the degradation of standard accuracy in the adversarial trained models and provide possible solutions.

In order to effectively prevent the occurrence of fire accidents, this paper proposes an improved algorithm for automatic identification of laboratory fire smoke based on GMM+YOLOv3. Firstly, the non-smoke area is removed by GMM foreground extraction to reduce the impact of less smoke in the early stage of fire and more irrelevant background on fire smoke recognition. Then, combined with the EfficientNet-B0 network, the feature extraction network of YOLOv3 is optimized to reduce the number of parameters, so that the model can be deployed in embedded devices to complete detection tasks. Secondly, in order to make it easier to detect small objects, the feature pooling fusion module is used in the model to improve detection accuracy. Finally, the model parameters are improved to enhance the training and detection effects of the network. Experiments have proved that compared with the traditional YOLOv3 target detection algorithm, this method reduces the number of parameters of the model, improves the calculation speed, and increases the accuracy rate by 5.9%, which ensures the accuracy of smoke detection in complex environments and effectively avoiding the fire accident.

Solar panel is an important tool to convert solar energy into electric energy, there are many defects in the production process. The defect area occupies a large size span in the whole image, in view of the above defect characteristics and the detection accuracy of traditional detection methods is not high. To solve the above problems, an improved Faster R-CNN based on bidirectional feature fusion module BiFPN is proposed, the multi-scale defect information is extracted by combining the strong semantic information of the high level feature map and the location information of the low level feature map. The experimental results show that the improved Faster R-CNN model has a detection mAP value of 88% for five kinds of defect samples, which is 11% higher than the traditional Faster R-CNN detection accuracy, which can better meet the needs of industrial practical applications.

In order to explore the feasibility of applying UAV remote sensing and object-oriented canopy extraction technology to forest areas with different canopy densities, 26 forest plots located in Gannan Plateau were selected and divided into two categories of coniferous forests with medium and high canopy densities. ESP tool was used to screen the optimal segmentation scale for multi-scale segmentation, and the feature space was optimized for object-oriented classification. The east-west and north-south crown amplitudes on quadrat scale were extracted in ArcGIS. The extracted crown amplitude was compared with the field measured data to evaluate the accuracy. The results were as follows: In UAV image extraction, the fitting values R2 and RMSE of east and west canopy width and measured values were 0.865 and 1.025m, respectively, and the fitting values R2 and RMSE of north and south canopy width were 0.867 and 1.235m, respectively. The R2 and RMSE of east and west canopy width with high canopy closure were 0.828 and 1.145m, respectively, and the R2 and RMSE of north and south canopy width were 0.787 and 1.523m, respectively. The extraction effect of the plots with medium canopy closure was better than that of the plots with high canopy closure, and both of the two canopy closure met the accuracy requirements. The results show that the automatic extraction method of canopy parameter information based on UAV remote sensing image is efficient and reliable, and has important practical significance in forest resources investigation.

Vision scheme based on the unmanned aerial vehicle (UAV) has gained widespread attention in recent years. Nevertheless, on account of the limited field-of-view of traditional images, it is difficult to apply to the fields that require comprehensive visual information. Therefore, a novel visual scheme based on the UAV platform with panoramic images is proposed in this paper, in which a platform consists of the panoramic cameras and the unmanned aerial vehicle are innovatively established and a scene recognition method based on discrete spherical image features are implemented. For the sake of reducing the distortion of panoramic image, we propose an icosahedron-based panoramic image representation for feature extraction, and then combined with the convolutional neural network and support vector machine, recognition task of the real image captured by the UAV platform are accomplished. Compared with the most widely used representation, namely the equirectangular projection, the proposed method can improve the recognition accuracy by 13.64% based on the Panoramic Scene dataset. Besides, our method can obtain a better performance even under the condition of large noise. Therefore, the proposed UAV-based panoramic scene recognition method can be applied to the fields that require comprehensive visual information effectively.

Aiming at the problems of difficult feature extraction, low detection accuracy and long detection time in traditional landslide detection methods, considering the diversity and complexity of landslide and its background environment, a landslide detection method based on mask region-based convolutional networks (Mask R-CNN) and Squeeze-andExcitation (SE) attention mechanism is proposed. The training set with 2050 samples and the test set with 600 samples were established for landslide detection, which verified the effectiveness of the improved method. The precision, recall and F1-score of the improved recognition are 89.2 %, 98.0 % and 93.4 %, respectively. Precision, recall, and F1-score all rose in comparison to the model before the improvement by 2.9%, 3.9%, and 3.4%, respectively. The findings demonstrate that the enhanced method can effectively increase the accuracy of landslide identification., and the landslide boundary segmentation results are closer to the real landslide boundary, so as to realize automatic and efficient landslide detection.

High frame rate and high grayscale are the main direction of development of the silicon-based micro-display. Therefore, based on the research of image dithering algorithm and multi-frame fusion scanning strategy, a high frame rate scanning strategy is proposed. Under the conditions of frame rate of 200Hz, data transmission bit width of 8 bits, grayscale of 256, and resolution of 1024×768, this strategy can make data transmission efficiency reach 100% and data transmission frequency reach 52.429MHz, which can support high frame rate display with limited bandwidth. Then use diffierent scanning strategies to process images. By the analysis of images, grayscale distribution histogram and grayscale test images, it shows that after the procession of diffierent scanning strategies, the grayscale linearity of images is good, the phenomenon of image distortion is not obvious. It proves that the strategy can meet the needs of silicon-based micro-display.

When verifying the metering chip, it is necessary to generate signal source excitation and write test cases, and build a simulation verification environment. The main content is the drive of incentives, the detection of output results, and the statistics of errors and data. In this paper, a signal source gain correction method is proposed, which realizes the simultaneous processing of the ratio difference between the effective value and the power value, reduces the number of ratio difference configurations in the calculation, reduces the deviation caused by the ratio difference configuration, and reduces the complexity of chip verification. And the automatic configuration of signal source gain is realized by software, which verifies the feasibility of the scheme. The experiment shows that the scheme can significantly improve the stability of signal source excitation in chip verification, reduce errors, and easily generate multiple signal source excitation through configuration.

UHV transmission lines have problems such as high impedance and low loss. This paper mainly introduces the monitoring method of UHV transmission and distribution line based on satellite remote sensing technology, aiming at avoiding or reducing property losses caused by line problems and maintaining people's normal life. In this paper, the experimental system construction and quantitative analysis methods are mainly used, and the results of two kinds of line detection are compared by comparison. The experimental data shows that in different types of fault lines, the test time of fault is within 0.1s.

As a popular task in remote sensing, road extraction has been widely concerned and applied by researchers, especially by using deep learning methods. However, many methods ignore the properties of roads in remote sensing images, which have long-range structures or discrete distributions. Therefore, this paper designs a network(MTP-Net) based on strip pooling and multi-scale spatial pooling. This network uses the ResNet50 network as the encoder to achieve feature extraction and ensures the overall connectivity and edge details of roads by the multi-type pooling module including strip pooling and multi-scale spatial pooling. The MTP-Net was tested on the Massachusetts Roads Dataset, the F1-score and IoU(intersection ratio) reached 72.24% and 56.54%, respectively. Compared with the mainstream methods such as UNet and deeplabV3+, the experiment shows that the MTP-Net is superior to the comparison model and has good results in road extraction.

Existing multi-person pose estimation methods show a tendency for the parameters in the model to become increasingly large in order to improve generalisation performance, requiring huge computational resources. Therefore, we propose a lightweight multi-person pose estimation method called PKDHP (Process Knowledge Distillation for Human Pose estimation), which applies knowledge distillation to pose estimation. PKDHP treats the training process of student models in knowledge distillation as a human learning process, specifically, PKDHP proposes the CABF module to replace the ABF module in ReviewKD aiming to guide the lower level feature through the model's higher level feature, and proposes the Transfer knowledge distillation method to force the student model to imitate the process of feature transfer from the teacher model, thus further simulating the human learning style. Compared with other knowledge distillation methods, we use the same model parameters on two multi-person pose estimation datasets (COCO and MPII) to achieve higher performance.

Due to interference in the transmission of external equipment, images can suffer from varying noise concentrations. As an immediate and practical step to reduce the impact of noise on an image and to improve its quality and visual presentation, filtering is the best method. The paper examines how traditional filtering methods restore a digital image under salt and pepper and Gaussian noise interference. Also included in this paper is an analysis of the advantages and challenges of traditional noise filtering, as well as suggestions for a filtering method for a variety of noise levels. Lastly, the effects of filtered noise pictures are evaluated through extensive experimental simulations, providing substantial assistance for related research.

The traditional dimensional measurement of gate tile steel back is still under the status quo of relying on manual work, which is not conducive to industrialization and automated production. In this paper, we rely on machine vision and image feature extraction to achieve contactless measurement of gate tile backs, preprocess the acquired images, binarization operation, edge detection using Canny operator, and finally obtain the product length and width parameters, square hole length and width parameters, central round hole diameter parameters using Hough transform and least squares method. The experimental results show that the method can accurately measure the index parameters of the gate tile steel backing products, and the detection error is within ±2mm. The method reduces the influence of human factors and environmental factors on the inspection of the gate tile steel backing, improves the measurement accuracy and speed of the quality inspection of the gate tile steel backing, and promotes the automation of the production of the gate tile steel backing to a certain extent.

The two-stroke low-speed diesel engine has complex vibration, and the conventional empirical mode decomposition (EMD) is not effective in its vibration signal processing. Thus, modified ensemble empirical mode decomposition (MEEMD) that a new method was employed to decompose the signal of the low-speed diesel engine block vibration. At first, a tri-axial accelerometer was applied to measure block vibration of the engine. After that, root mean square (RMS) was used to analyze the vibration intensity in three axes. Finally, the MEEMD method were performed for the signal in x-axis. The results show that the block vibration intensity in the x-axis is the largest. And the MEEMD method can suppress the modal mixture compared to EMD which help to better identify block vibration excitation.

The pancreas is located deep in the abdominal cavity, and its structure and adjacent relationship are complex. It is very difficult to treat it accurately. In order to solve the problem of automatic segmentation of pancreatic tissue in CT images, we apply the multi-scale idea of convolution neural network to Transformer, and propose a Multi-Scale Swin Transformer and Complementary Self-Attention Fusion Network for Pancreas Segmentation. Specifically, the multi-scale Swin Transformer module constructs different receptive fields through different window sizes to obtain multi-scale information; the different features of the encoder and decoder are effectively fused through a complementary self-attention fusion module. By comparing experimental evaluations on the NIH-TCIA dataset, our method improves Dice, sensitivity, and IOU by 3.9%, 6.4%, and 5.3% respectively compared to the baseline, which outperforms current state-of-the-art medical image segmentation methods.

For mechanical dimension detection of Printed Circuit Board (PCB), an image feature recognition method based on location relationship was proposed. Select a standard PCB standard graph as the standard graph, establish a template for image matching in the standard graph, and determine the measurement area, measurement type, standard size and tolerance. Through template matching of test points in PCB test, comparing different types of test characteristics, selecting appropriate image feature extraction and identification, comparing standard size and tolerance, so as to determine the quality of PCB. This paper introduces a fast and efficient automatic test technology based on PCB test requirement.

With the application of a new generation of information technology in the field of industry and manufacturing, the era of big data-driven manufacturing is coming. However, as the most basic component of the digital manufacturing workshop, the distribution of manufacturing resources such as personnel, equipment and products in the workshop is highly uncertain, so the management of manufacturing resources has become a major difficulty in the safe production link of the digital manufacturing workshop. Compared with the traditional target detection algorithm, this paper greatly reduces the detection omission problem due to occlusion through multi-field fusion, and improves the positioning accuracy of workshop manufacturing resources through absolute coordinate mapping and data fusion. We train the model from scratch, achieved 46.5 box AP and the max positioning deviation is 0.81m on NJUST workshop dataset.

Based on the research of single-channel reluctance resolver, the signal winding distribution function of a new type of coexcitation axial reluctance resolver(CARR) is summarized. The self-inductance and mutual inductance of the excitation windings and signal windings belong to the coarse machine and precise machine under the even pole pairs are studied theoretically. The relationship between inductance of different windings and structural parameters is obtained. The correctness of the inductance distribution law is verified when the number of pole pairs is even by comparing the finite element simulation method with the theoretical research method. It provides the basis for the theoretical design of this resolver. The self-inductance and mutual inductance of 15 pole pairs CARR are measured and compared with the finite element simulation results of inductance. Then the inductance distribution law is summarized when the pole number is odd, which is convenient for the subsequent design and optimization of CARR.

This paper presents a new type of co-excitation axial reluctance resolver(CARR), and deduces the distribution functions of signal windings in the winding mode of equal turns. According to the flux change law of the upper and lower teeth of the stator, the expression of axial effective coupling length of rotor is obtained. The self-inductance and mutual inductance of two-phase precise machine signal windings and excitation windings are theoretically studied by winding function method. The inductance variation law of each winding of CARR is obtained under the condition of odd pole pairs and even pole pairs. It provides the basis for the theoretical design and optimization of this kind of resolver, and is beneficial to improve the measurement precision of precise machine signal output. Then, the finite element method is used to compare and analyze the inductance variation law of each winding obtained by analytical method, which verifies the correctness of the theoretical research on the inductance of each winding in the winding mode of equal turns.

As the remote sensing image information rapidly becomes abundant, it is a challenge for the detection of tiny targets with dense distribution. Therefore, a multi-scale rotating object detection model based on the improved YOLOv5 is proposed in this paper. Firstly, because of adding a prediction feature layer to the network, the detection precision of tiny targets has substantially increased. Secondly, the loss between rotation anchors can be fitted at a high precision due to the compressed loss function aiming at the calculation of the IoU loss of rotation is proposed. Finally, a hybrid prior bounding box strategy is applied in the feature prediction layer to suit the targets to be detected in different sizes. Experiments conducted on the DOTA dataset indicates that this method significantly exceeded the original YOLOv5. It has extraordinary performances for the task of object detection in optical remote sensing image fields.

Target detection is one of the hot research issues in the field of computer vision in recent years. Large scale convolution networks can effectively improve the accuracy of target detection, but they are not suitable for application scenarios with limited computing and storage capacity. In order to solve the above problems, this paper proposes a lightweight target detection network based on improved YOLOv5 - CS-YOLOv5 network. This network uses a lightweight network ShuffleNetv2 to reconstruct backbone, effectively reducing the network complexity, and adds a CA attention module to the feature extraction network to enhance the network feature expression ability, and solves the problem of neuron deactivation through H-Swish activation function. The final model size was compressed to 1.64MB, and the mAP50 on the COCO test set reached 47.6%, reducing the model size by 13.7% and improving the detection accuracy by 1.9% compared with the original YOLOv5 model.

Aiming at the problems of low measurement efficiency, large error and inaccurate roundness evaluation of bearing roller size, a roundness and cylindricity detection system based on machine vision is established. The calibration plate is used to complete the size calibration. The image is collected by rotating the roller at a fixed angle, and the water level is corrected. A third-order gray moment subpixel detection algorithm based on improved Canny operator is proposed to locate the roller edge subpixel. Then the improved minimum zone circle (MZC) is used to evaluate the error and the non-contact and the non-contact online measurement of the roundness and cylindricity of bearing rollers is realized. The test results show that the average roundness error obtained by this method is 10.2 μm. The average cylindricity error is 13.5 μm. The measurement result and accuracy are 10 μm. Compared with the three-dimensional contour scanner, the detection accuracy is improved by 10%, and the measurement error is controlled at 12 μm. The sub-pixel precision obtained in this study can achieve high-precision measurement of roundness and cylindricity, which provides a reference for the following such issues.

Automatic segmentation of the liver from the patients CT is of great significance for the diagnosis of liver diseases. Because the use of bottom-up feature fusion in U-Net ignores the importance of low-level features, resulting in poor network segmentation performance, and the similarity of gray values between liver and adjacent organs and tissues, making some small detail features difficult to pay attention to. In this work, we propose a network called DMU-Net. DMU-Net replaces the original convolution module with dual-channel multi-scale convolution to obtain richer semantic information. In addition, the attention mechanism is added between the encoder and decoder to increase the feature extraction ability of the model for small targets. The results on LITS dataset show that the DICE coefficient and IOU of the proposed model get 97.46% and 93.27% separately. Compared with the classic U-Net network and MultiResUNet, the model can segment more accurate liver regions and achieve better segmentation results.

Oil immersed shunt reactor is one of the most important reactive power compensation equipment in Ultra high voltage (UHV) long-distance power transmission and transformation system. When the oil immersed shunt reactor has potential defects, the vibration signal during operation will be different from that under normal condition, which is shown as different defect signals such as noise and abnormal temperature. Compared with traditional vibration sensor detection, acoustic detection and infrared detection realize non-contact detection, which is safer and does not require complex vibration sensor layout. If the two are combined, the detection efficiency can be greatly improved.In this paper, the acoustic characteristics and thermal distribution characteristics of oil immersed shunt reactors under different vibration modes are studied. Based on the beam forming sound source space positioning calculation method, after the integration of voiceprint and infrared image is completed, the acoustic imaging and infrared detection technology of reactor vibration is proposed to realize the simultaneous monitoring of reactor sound temperature state information, and improve the accuracy and rationality of defect identification. In addition, the design scheme of a prototype for on-site acoustic infrared detection of reactor vibration is proposed. The related results of this paper can provide technical support for the application of UHV oil immersed shunt reactor state assessment and diagnosis, and improve the level of equipment operation and maintenance.

It is currently unclear how to calculate the distance between two Pythagorean fuzzy sets. Several techniques have been proposed. However, not all available approaches can accurately exhibit differences between Pythagorean fuzzy sets and fulfill the axiom of measurement characteristic. And some other methods ignore the relation between the three variables of the Pythagorean fuzzy set. A new direction of estimating distance is presented to address the issue. Then some numerical examples are provided to demonstrate that the way of measuring distances avoids producing counter-intuitive and irrational conclusions and is more effective and reasonable than other similar methods. Furthermore, the suggested method of measuring distances between PFSs is evaluated in a real-world application scenario and compared to earlier approaches to illustrate its superiority and efficiency. At the same time, the viability of the suggested method in dealing with uncertainties in practice is demonstrated.

In view of the traditional aerial image of UAV, the target detection method has many shortcomings, such as inaccurate detection and positioning, low detection accuracy and so on. Based on the original YOLOv5 framework, this paper proposes a new target recognition algorithm and an improved feature extraction method, and uses ultra-lightweight convolution neural network MobileV3 to replace the main feature extraction network of YOLOv5 network; In order to improve the attention of the model to other regions, the attention mechanism GAM Attention module is introduced. This mechanism embeds the location information into the channel attention, realizes multi-scale processing and feature fusion, and makes the model have higher detection accuracy. The experiment shows that the size of the improved lightweight network model is only 81.7% of the original YOLOv5 model, and the mAP on the data set reaches 96.8%. This method greatly reduces the model parameters and the amount of computation on the basis of ensuring the detection accuracy, and also greatly improves the detection speed and accuracy. Therefore, the method proposed in this paper can provide strong support for the application of UAV target detection.

Underwater pipeline transportation, as an important means of natural gas transportation, will cause great economic loss and environmental pollution in case of leakage damage. Aiming at improving the accuracy of pipeline leakage monitoring research, a method was proposed to monitor the leakage process of underwater natural gas pipelines using distributed optical fiber acoustic sensing technology. In this paper, a processing algorithm combining empirical modal decomposition method and wavelet decomposition reconstruction is used to extract the signal frequency domain features. The experimental results show that the frequency domain amplitude of the vibration signal gradually grows as the leakage orifice diameter and internal pipe pressure increase, and the standard deviation of the vibration signal of the pipeline exhibits a quadratic fitting relationship with the size of the leak aperture. The method proposed in this paper has a minimum leak aperture identification error of 7.2%, which can effectively improve the pipeline leak monitoring accuracy.

With the improvement of the intelligence degree of textile and garment industry, with the help of advanced information technology, using computer image recognition method to realize automatic fabric recognition, can improve the recognition accuracy and production efficiency to a great extent. At present, the common fabric recognition algorithms do not combine the basic properties of fabric, only consider its image features, so the recognition accuracy is low. In this paper, on the basis of predecessors' research put forward a kind of fabric recognition method in combination with the fabric properties, with the commonly used 12 kinds of natural fiber fabrics as samples, geometric measurement of fabric image was established, the quantitative analysis of impact resilience and drape of fabrics attribute two factors, and the fabric properties of parametric modeling, get the fabric properties of geometric measurement. The image measurement data were normalized and constructed by polynomial regression. The convolu-tional neural network (CNN) is trained to extract features adaptively according to image features. The fabric recognition model is constructed to learn the parameters of two fabric attributes and realize automatic fabric recognition. It can effectively improve the accuracy and efficiency of cloth recognition and meet the needs of practical applications.

The digital construction process of power grid often uses image intelligent recognition technology in the transmission and substation system for equipment environment monitoring, unmanned inspection and other scenarios, and this process generates massive data, and the existing centralized data processing model is limited by the lack of storage resources and computing power, and cannot meet the massive data processing requirements and frequent interaction needs. In this regard, this paper proposes a framework for the application of image recognition data of side electrical equipment based on cloudside - end collaboration, and builds an online intelligent diagnosis and analysis model of electrical equipment based on image recognition algorithm, so as to achieve real-time tracking of the operation status of electrical equipment. To further verify the stability and safety of the built model, Matlab simulation software is applied to conduct a targeted case analysis of the model, and the superiority of the proposed cloud-side synergy in the image recognition model of electrical equipment is verified through the comparative analysis of data synergy efficiency curves, which proves the feasibility of the proposed design scheme.

This paper is mainly implemented through neural network algorithms, using Python language, for identifying, processing the speech of children with autism, and then comparing with the speech database established in advance to obtain the final speech results, on this premise, to achieve human-computer interaction, to achieve the purpose of practicing dialogue with children with autism or guiding them to speak, in order to restore the language function of children with autism.

This paper mainly introduces the application of Digital Image Correlation (DIC) method in the structural safety monitoring of high pile wharf, and describes the basic working principle, equipment composition and monitoring effect of the method in the laboratory environment. According to the test results, the real-time monitoring of the state of the large high-pile wharf structure can be achieved by the binocular vision based on real-time sensing technology, and the key index of the monitoring -- the time-history curve of local displacement has a good correlation with the actual displacement. Furthermore, the results lay a foundation for the landing application of the technology, and provide data guarantee for the real-time monitoring of structural safety of wharfs with longer service time.

The current traditional deep learning-based sound recognition algorithm achieves the classification and recognition of sound by constructing MLP models, which leads to a large computational effort and low sound recognition rate due to the lack of dimensionality reduction processing of sound signal features. In this regard, we propose the study of sound recognition algorithm for power plant equipment by fusing MFCC and IMFCC features. Pre-processing such as pre-emphasis, normalization and framing of the sound signal is performed, and a decentralized approach is used to realize the dimensionality reduction of the sound signal features and achieve the recognition of power plant equipment sounds. In the experiments, the average recognition rate of the proposed algorithm is verified. The analysis of the experimental results shows that the sound recognition algorithm constructed by the proposed method has a high average recognition rate and a good recognition effect.

Gait recognition has achieved high recognition accuracy under normal walking conditions, but in the case of occlusion, the recognition efficiency needs to be improved. To solve this problem, this paper improves the Gaitset network: change the primary feature extraction network into dynamic region-aware convolution, and dynamically generate corresponding filters according to the feature distribution of different regions of the gait image to extract features, so as to enhance the representation ability of convolution; GEI is added to the MGP module to make up for the lack of global information in the deep network. The experimental results on CASIA-B dataset show that the gait recognition rate of this algorithm under NM, BG and CL conditions is 3.4%, 5.0% and 6.9% higher than that of GaitSet algorithm. It is proved that this experiment has certain advantages for gait recognition under occlusion.

Binocular vision-based target detection is one of the hot topics in computer vision, where the technique aims to detect and localize target objects in images. The technology has applications in fields such as autonomous driving, video surveillance, and UAV flight control. In recent years, with the development of deep learning techniques, its speed, accuracy, and robustness have led to its widespread use in various research areas. This paper first lists the history of the development of target detection techniques, then introduces two target detection methods for binocular vision, and finally suggests possible improvements and development trends. Through summary and analysis, the aim is to provide a reference for work related to conducting binocular vision target detection.

Transformer-based object tracking has become mainstream due to its high tracking performance at present. Most transformer-based trackers still classify the foreground and background and then calculate the bounding box by regression. However, we find that the classification confidence of transformer tracking is inaccurate. This leads to the bounding box with the highest confidence is not the closest to the ground truth. To address the problem, we propose a robust visual tracking approach with confidence correction using log-polar mapping. By correcting the confidence, the transformer tracking can obtain the best bounding box and improve the accuracy of the tracker. Extensive experiments on OTB2015, UAV123, and UAV20L show that our tracker is superior to the state-of-the-art trackers.

Visual-inertial odometry is widely used in robot motion estimation due to its excellent performance in dealing with scale ambiguity. In this paper, an end-to-end monocular visual inertial odometry is proposed, and its innovation lies in its feature selectivity to extract key features to estimate motion, hence we call it SF-VIO. The most significant difference between our proposed SF-VIO and other methods is that SF-VIO uses key features to estimate robot motion more accurately after feature selection. The feature selection is divided into two steps. Firstly, SF-VIO implements the Attention Mechanism in the image process module Locator to filter useless image information. Secondly, SF-VIO implements a soft fusion mechanism in the feature fusion module Fusion, which effectively integrates visual features and inertial features. The experimental results on the KITTI dataset show that the SF-VIO algorithm outperforms similar methods such as DeepVO and ESP-VO in accuracy.

Nonnegative matrix factorization (NMF) is a feature learning method that can achieve nonlinear dimensional approximate reduction with strong interpretation, and it is widely used in the field of tumor recognition. The objective function of the traditional NMF model is based on the Euclidean distance metric, and the performance of the model is easily affected by the noise. Moreover, traditional NMF is an unsupervised feature learning method that does not use the label information of the data. However, it would cause a waste of information without using label information and cannot learn the discriminative features in the data. Therefore, the supervised nonnegative matrix factorization model with fused correntropy (FCSNMF) is proposed in this paper. The FCSNMF model alleviates the effect of noise in the experimental data by fusing the Euclidean distance metric and the maximum correntropy metric. In addition, the label consistency regularization term is skillfully chosen to utilize the label information of the data to obtain discriminative features. The effectiveness of the FCSNMF model is verified by applying it to a gene expression profile dataset for tumor recognition.

In this work, a lightweight object detection network based on YOLOv5 is proposed to solve the detection problem of nonmotor vehicle helmet wearing on the road. By replacing the original CSPDarknet53 backbone network with MobileNetv3, the number of network model parameters is reduced and the computing speed is increased. With further introduction of the criss-cross Attention, it improves the accuracy of the simplified network and improves the detection reliability. Compared with YOLOv5 network, the detection efficiency is improved 140%(FPS) at the cost of reducing 17.1%mAP, hence the task of helmet detection on the road can run in real time on the edge computing equipment.

As the transportation industry continues to grow and traffic conditions become increasingly complex, ensuring driving safety has become a pressing issue of concern. Traffic sign recognition can offer real-time evaluation of road conditions and driving safety alerts for vehicles and ultimately reduce traffic safety accidents. This study proposes a class-imbalanced traffic sign recognition method based on improved YOLOv7. Firstly, the Coordinate Attention is incorporated into the YOLOv7 backbone to enhance the representation and localization ability of the target to be tested. Then, the larger feature layer is introduced to improve the recognition ability. Finally, the class imbalance problem in the dataset is addressed by using Cost-sensitive softmax cross-entropy loss. The improved YOLOv7 network has been shown through experiments to enhance the detection capability of traffic signs effectively, reaching 90.54% mAP and 86.78% F1-Score.

As a biometric recognition technology with the advantages of long-distance and anti-camouflage, gait recognition has great application value in public security criminal investigation, security monitoring and other fields. Due to the influence of external factors such as viewing angle, clothing and light on the appearance of gait sequence, most of the current gait recognition algorithms have low accuracy and cannot be applied on a large scale. Aiming at the view angle problem in gait recognition, we proposed a method of constructing full-view gait gallery set based on dual Kinect cameras and edge point cloud registration. Then, we used the RGB camera to collect gait sequence from any angle of view (including depression angle), and designed a gait feature matching network called GaitCrossNet, which is used to match the RGB gait contour sequence to the sequence projected from point cloud sets. The experimental results showed that the average recognition accuracy of the algorithm proposed in this paper can reach 85.2%, which has obvious advantages compared with the existing recognition schemes.

With the development of power system, the technology of power well is facing many problems. How to not only detect the power well information, but also characterize the measured data in an easy-to-explain way, so that the staff can understand and analyze the power well casing, is of great significance to the detection of pipeline damage and the development of the power industry. Nowadays, the development of informatioa n processing and microelectronics technology, especially the rapid development of image processing technology, provides a more theoretical basis for the realization of power well detection technology, and expands more possibilities for the study of detection data. A clutter suppression method based on fine-grained optimization decomposition of the convolution feature is proposed. A noise constraint is added into the non-negative matrix factorization, and a base matrix and a coefficient matrix to be factorized are assumed to be non-negative exponential distributions in probability. The non-negative matrix factorization is carried out by using a convolutional neural network to obtain the approximate inference of the base matrix and the coefficient matrix. The selection of parameter values and the input of horizontal gradients are determined. The low-rank matrix representation of clutter components is obtained so as to separate the clutter from the image. The experimental process is compared with the simulation data, which verifies that the model has good robustness for the image recognition in the power well.

Aiming at the problem that the traditional convolutional neural network is easy to lose a lot of useful information in the expression feature extraction stage, and cannot extract the highly discriminative expression feature, which leads to the low expression recognition rate, a method of facial expression recognition based on multi-scale feature fusion attention mechanism is proposed. First, ResNet50 is used to extract the convolutional feature. In order to avoid the loss of expression feature information, multi-scale feature fusion is performed on the output feature images of different convolution layers in the network, and more abundant expression feature information is extracted while introducing context information; In order to focus on the key expression features, an improved hybrid attention mechanism is introduced into the network to learn the weight information from channel and spatial domain, obtain the attention map, enhance the expression ability of features, and suppress the impact of redundant information; In order to further improve the discriminability of the extracted expression features, pyramid convolution is introduced to improve the residual network structure. The experimental results show that the average accuracy of the model on the RAF-DB dataset is 83.67%.

Image recognition technology is one of the typical application 2 3scenarios of artificial intelligence in the field of power production technology. It is of great significance to carry out the research and application of image recognition technology in the intelligent patrol inspection of substations for promoting the construction of substation intelligent stations. In order to reduce the pressure of operation and maintenance of secondary panels and switch cabinets in the substation, and solve the problems of multiple procedures, long cycle, low efficiency and error-prone in leak detection, a method of intelligent detection of the status of substation panels and cabinets using image recognition technology is proposed. In practical applications, the identification and analysis of the results of the substation inspection robot and video camera can effectively improve the detection efficiency, reduce the operation and maintenance costs of power equipment, and realize paperless operation, which is of great significance to the realization of intelligent inspection of the substation.

Biological features are widely used for person identification, lip prints have been proved to be a unique and permanent part of the body. Therefore, lip print recognition can be considered as an effective mean to confirm individuals. This paper developed an architecture with convolutional spiking neural network (CSNN) for lip print recognition. Spiking neural networks (SNN) have the potential to reduce power consumption comparing to traditional artificial neural networks (ANN). The leaky-integrate-and-fire (LIF) neuron model and gradient surrogate method were used in the proposed CSNN. This work provided a novel strategy for lip print recognition.

In order to reasonably develop and utilize groundwater resources, this paper proposes to use multi-source data to identify the location of groundwater wells in the Ningzhen Mountain area. An index system of influencing factors for groundwater supply identification of bedrock is established by analyzing relevant factors affecting groundwater occurrence. AHP is used to assign index weights and verify the correctness of weights. Then, combined with industry experience and expert knowledge rules, the correctness of the identification is judged by using bedrock geological maps and profile maps to identify the water storage structure and contour data to identify the terrain to obtain the water control points in the study area. The identified water storage structure and water control points are spatially superimposed to obtain the possible well location. Finally, combined with hydrological data, comprehensive analysis is carried out to determine the strength of the water-rich degree of the structure and then determine the well layout point. Comparing the experimental results with the collected data of the experimental area, the correctness of the scheme is proved. Compared with the traditional method of exploiting groundwater by relying on manual experience, this method of scientific exploitation of groundwater can more intuitively display the location of suitable wells on the map, with a high success rate of exploitation. In addition, it can avoid the environmental problems caused by blind exploitation.

Analyzing facial micro-expressions based on computer vision has become an active research area in recent years. Subtle changes and short duration of micro-expressions bring great challenges to this issue, and extracting effective spatiotemporal features is the key to improving recognition performance. This paper proposes an approach for micro-expression recognition, Spatiotemporal Local Gabor Binary Patterns on Three Orthogonal Planes (STLGBP-TOP), by combining spatiotemporal Gabor filtering and LBP-TOP. In this approach, LBP-TOP is employed to encode local dynamic features of micro-expression videos on top of spatiotemporal Gabor filtering. Experimental results on CASME and CASME2 datasets demonstrate the effectiveness of our method.

Existing methods related to event-based SLAM mostly realize front-end visual odometry function in visual navigation and it is difficult to handle loop closure task directly with accumulated events, in which case a complete SLAM framework is needed to be accomplished. This paper presents a solution to the problem of loop closure detection in real time event-based visual odometry with assistance of reconstructed intensity image. The proposed method selects keyframe in frontend and detects loop closure using calculated bag-of-words from intensity image. While loop closure is detected, the photometric bundle adjustment method is implemented to refine global camera pose and global 3D depth points constraints. The proposed method has been tested in MVSEC event-based stereo datasets with loop information, and successfully detects and corrects loop closure. Experiments have shown that the proposed method can successfully reduce about 50% error triggered by visual odometry in real-time trajectory estimation.

To address the problem that the direction of pipe cracks is difficult to detect, a crack direction recognition method based on prototype learning is proposed with a prototype network as a framework. First, through the convolutional layer of the prototype network, shallow features of crack directions are extracted to improve the generalization ability of the model on the data set of this paper. then, by improving the High-Resolution Network and introducing a location self-attention mechanism, and combined with a migration training method for the data set of this paper, a category that can accurately reflect the crack directions is constructed prototype learning mechanism. Finally, pattern recognition is performed by the metric classification methods, the effective classification of crack direction under small sample condition is achieved. The experimental results show that the recognition accuracy of the crack direction recognition method based on prototype learning can reach 99.2% with the sample parameters unchanged.

Multimodal Sentiment Analysis (MSA) has achieved substantial progress as an open and practical research field. Several methods in MSA focus on exploring the methods for multimodal data fusion to improve the model performance. However, the heterogeneity gaps pose a significant challenge against the fusion interactions of the multimodal data. This work uses a new Multimodal Sentiment Analysis (MSA) model, the adversarial invariant-specific representations fusion (AISRF) network. AISRF is proposed to achieve modality-invariant representations by narrowing the distribution gaps among different modalities. At the same time, the integrity of the modality-specific representation is maintained. First, the heterogeneity gaps among the modalities are reduced by invoking an adversarial encoder-regressor, and thus the modality-invariant representations are obtained. Second, the decoders are employed to reconstruct the modality-invariant representations and obtain the modality-specific representations. Finally, the cross-modal attention method has been used to perform the cross-modal interactions on the invariant-specific representations from different modalities to perform efficient multimodal fusion. Experiments for comparison with other baseline models have been performed on the prevailing benchmark datasets, viz., CMU-MOSI and CMU-MOSEI, and the results demonstrate that the AISRF model is superior to the baseline models in the multiple evaluation indices.

Semantic segmentation for autonomous driving has high-speed requirements, but the current semantic segmentation model requires a lot of computing resources and is very time-consuming, which makes it less suitable for autonomous driving. To solve this problem, we propose a fast real-time semantic segmentation network, which has a minimal amount of computing and can run in real-time, so it is very suitable for the self-driving field. The framework contains two essential modules, FPN of images and shallow Backbone. It ensures the efficiency of image multi-scale features and feature extraction, respectively. Experimental results on cityscapes show the effectiveness of our method.

Aiming at the problem of crop pests and diseases in agricultural production, this paper implements an identification algorithm of crop pests and diseases based on improved DenseNet model to achieve real-time detection of crop pests and diseases and early warning. Based on the DenseNet neural network, the crop dataset is first subjected to image processing and augmentation such as Canny edge detection, flipping, convolution and blurring, and the resulting dataset is used to train the DenseNet model. Moreover, the innovative addition of a pooling layer and a fully-connected layer to the DenseNet allows the model to obtain accurate identification results of crop health conditions with corresponding probabilities. The algorithm was tested on Plant Pathology 2020 - FGVC7, and the experimental results show that it is faster than traditional recognition algorithms, with a correct recognition rate of 96.7%, which can quickly and accurately diagnose crop pests and diseases and effectively improve crop yield and quality.

The use of eye movements for identifying intrinsic human psychological changes is a hot topic, and the identification of eye states in specific scenarios is the basis for analyzing the test taker's instantaneous psychological response state and thus for testing implicit driving quality deficits. Infrared eye images are difficult to visualize eye movements, and traditional eye image processing results are not suitable for eye state analysis. In order to make the eye images show the eye movements, a feature extraction method based on contours and key point is proposed for infrared eye images. To extract the eye contour features, the iris and eyelid contours are extracted based on the variation of gray values between the points located on the boundary and their surroundings; for the extraction of eye key point features, a network design with a combination of dlib and residual structure is used to extract the pupil and eyelid key points, respectively. The experimental results show that the method can represent the relative positions of the pupil, iris and eyelid separately with little computational effort, which can intuitively show the eye movement and thus reduce the stress of subsequent eye state recognition.

With the continuous rise of broadband radar in recent years, the information acquisition of space targets has gradually expanded from the original RCS, orbit, one-dimensional range profile and other one-dimensional information to twodimensional image information. The traditional recognition and matching of space targets based on orbit information plays an important role in certain scenes, but for some orbit change targets, especially for orbit change outside the detectable range, To a large extent, it is necessary to recognize according to the shape of the target, and high-resolution ISAR images are helpful to recognize and judge the orbit changing target. This paper presents a spatial target recognition method of ISAR Image Based on principal component analysis (PCA). The recognition effects of training and testing in different imaging time periods are studied successively, and compared under different recognition methods. Based on the constructed target database, the principal component analysis method has a better recognition effect.

To improve the real-time performance of radiation source signal time-frequency analysis, a fast radiation source signal time-frequency feature extraction method based on low-order cyclic statistics and Choi-Williams transform is proposed. First, the low-order cyclic statistics of the radiation source signal are calculated. Second, the IFFT is used to recover the cyclic domain signal, and the number of sampling points of the radiation source signal is adaptively controlled by the cyclic mean and IFFT point number. Finally, the recovered signal is used for CWD time-frequency analysis. In order to highlight the time-frequency image features, the geometric features of the signal modulation type are extracted using binary image transformation. Simulation analysis shows that the proposed time-frequency preprocessing method can significantly reduce the computational complexity while maintaining the accuracy of time-frequency features and improving real-time performance, providing effective support for subsequent sorting and recognition and parameter estimation work.

Audio type recognition based on audio feature value matters a lot to audio indexing. To attain accurate results, an effective calculation or extraction of audio features is paramount. This paper uses Discrete Wavelet Transform (DWT) on audio signals to obtain the statistical eigenvalues of audio by means of a comprehensive eigenvalue calculation method combining wavelet transform and frequency domain. Additionally, the Support Vector Machine (SVM) Model was used to construct different audio differentiation templates, thus enabling the recognition of various types of audio. The experimental results have revealed that the recognition method of the combined feature and Support Vector Machine model can obtain relatively high accuracy and demonstrate considerable practical application value.

Aiming at the fundamental role of railway wagon number recognition in railway freight management and railway wagon inspection, a railway wagon number recognition method based on Combined Grey Wolf Optimizer Support Vector Machine (C-GWO-SVM) is proposed. Aiming at the confusing numbers and letters in the railway wagon number dataset, the algorithm in this article first uses a set of GWO-SVM to perform multi-classification processing on the railway wagon number characters, and divides the numbers and letters in the dataset into easily distinguishable number and letter characters, and easily confusing number and letter characters, with a classification accuracy of 98.875%; Then two sets of GWO-SVM are used to classify and recognize the numbers and letters in the railway wagon number characters, with classification accuracy of 99.70% and 99.99%, respectively. The experimental results show that compared with Sparrow Search Algorithm (SSA) to optimize SVM multi-classification algorithm, the GWO-SVM algorithm has shorter parameter optimization time, higher recognition accuracy and faster recognition speed in the application of railway wagon number and letter characters recognition.

Recently, machine learning attracts more attention on autism spectrum disorder (ASD) identification based on resting-state functional magnetic resonance imaging (rs-fMRI). Most studies on ASD identification use rs-fMRI data from multiple imaging sites to increase sample size, but they suffer from the data heterogeneity among sites. Besides, most ASD identification studies use features simply extracted from brain connections, ignoring the topological structure of brain. To address these issues, we propose a brain graph synthesis model based on generative adversarial network (GAN), which transforms data from the source domain to the target domain, solving the data heterogeneity among sites. Specifically, the generator and discriminator are designed for graph structured data and propose topological losses to construct the cycle consistency loss for maintaining the original topological structure of the reconstructed brain graph. We carry out experiments on six tasks using the open database Autism Brain Imaging Data Exchange (ABIDE) for ASD identification. Experimental findings demonstrate that our model can solve the problem of data heterogeneity effectively and achieve satisfying performance on multi-site ASD identification.

Pedestrian attribute recognition has attracted extensive attention in surveillance scenarios. Many researchers treat it as a multi-label classification problem. However, they ignore the relations among attributes. For example, long hair and female often appear together at the semantic level, and there exists an overlap between their attribute-related regions at the visual level, e.g., female-related regions including long-hair-related head area and skirt-related areas. There is a correspondence between the semantic level and the visual level. However, existing relation-based methods which explore relations among attributes mostly ignore the correspondence between semantic and visual levels in relations among attributes. Our main contribution lies in proposing a novel end-to-end trainable Attribute-related Graph Visual Dependencies (AGVD) framework that can learn the dependencies of attributes based on extracted attribute-related visual features. Moreover, we propose a Graph Fusion mechanism to guide the attribute-related graph generation and pruning via semantic relations innovatively. We conduct experiments on three large-scale pedestrian attribute datasets. The experiments on PETA, RAP, and PA-100k have demonstrated that the AGVD outperforms the previous state-of-the-art methods.

For the process of wave detection in sea area, the traditional buoy monitoring spatial distribution is uneven, and satellite remote sensing is time-consuming and unfavorable to the safety of dredging construction. In this paper, the video/image feature detection method based on visual information, using UAV to obtain video information of the sea environment and automatically extracting wave features through convolutional networks, it can achieve the purpose of inferring regional wave changes from a single image. The network model combines a lightweight Mobilenet-V2 deep convolutional network in the network architecture, based on deep separable convolution and superimposed residual connections, which has the advantages of fast speed and high accuracy, and can realize real-time computation at the shipboard terminal and the network requires less additional computational resources, which reduces the need for energy supply and facilitates the deployment of large-scale dredging projects. The experimental results show that the network identifies 89% of samples with wave height error less than 20% and 94% of periods, which has strong robustness and good practicality in the field of ship dredging engineering.

This paper describes a machine visual recognition method for the surface rust recognition of metals. The corrosion images of iron-based metals (HT200, Q345B, 45 steel and 304) under the specific conditions can be accurately recognized by the proposed method. Based on the rust recognition method, the corrosion area percentage (CAP) of each metal block is calculated. The CAP results indicate that HT200, Q345B and 45 steel are rapidly corroded (CAP over 86 % after 192 hours), but 304 is slightly corroded (CAP: 1.7 % after 192 hours) under the periodic immersion condition. The correlation analysis between CAP values and actual weigh loss values is carried out to evaluate the proposed method. The results show that there is a high correlation between CAP values and actual weight loss values of each metal block. Therefore, CAP values calculated from the rust recognition method can quantitatively identify the corrosion degree of the metal blocks. No matter what type of the iron-based materials is, the corrosion degree of metals can be assessed by the rust recognition method. Even the serious corroded metal blocks, the corrosion degree can be determined through the surface rust recognition.