This PDF file contains the front matter associated with SPIE Proceedings Volume 12509, including the Title Page, Copyright information, Table of Contents, and Conference Committee list.

In the past, video games were functionally limited to displaying monochrome pixels on an 8-bit screen. As the technology has evolved, this form of expression has been preserved for its unique nostalgic aesthetic and evolved into pixel art. Many excellent teams and independent developers have adopted the pixel art style and produced amazing works at low cost. In this paper, we develop a pixel art game based on human-computer interaction. The layered tetrad method is taken as the game design principle and realizes the interactions in this game. In terms of gameplay, this game will also take the knowledge in the introduction to gaming as the guiding theory, so that the works can bring better interactive experience to players. Throughout the game, players will experience immersive adventures, combat interactions based on mathematical models, and a well-acted story with a fragmented narrative style.

With the rapid development of information technology, online open courses are playing an increasingly important role in university education. By collecting the course resource information and excellent cases of teachers' online teaching in the online open course alliance platform of universities in the Guangdong-Hong Kong-Macao Greater Bay Area, the article conducts research from six dimensions: the number of course sources in the alliance platform, the classification of course subjects in the alliance platform, the number of courses offered by alliance institutions, the number of courses introduced by alliance institutions, the number of learners studying in alliance institutions and the choice of online teaching wisdom tools of teachers in alliance schools, and analyzes the construction and application status of online courses in the alliance platform. Finally, it puts forward the strategic measures to improve the level of online course construction of the alliance platform, and provides some reference suggestions for teachers to carry out online education and teaching.

Using learning analytics technology to mine online learning features can optimize the teaching process. On the basis of collecting students' online learning information, the similarity between features is firstly analyzed; secondly, machine learning algorithm is used to construct a student's learning performance prediction model, and the accuracy and K value of the model are analyzed; the teacher's teaching process reform puts forward the corresponding direction.

Gesture recognition can play a crucial role in addressing the issue of Human-computer interaction. In this paper, we proposed a vision-based Multi-input fusion deep network (MIFD-Net), which consists of Multilayer Perceptron (MLP) and Convolutional Neural Networks (CNN). MIFD-Net first processes hand keypoint data and gesture images using Euclidean distance normalization (ED-Normalization) and image segmentation technologies, respectively. Then, two kinds of data are simultaneously used as input to MIFD-Net. The experimental results show that the MIFD-Net achieves an average accuracy of 99.65% on the self-built dataset in this paper and 99.10% on the NUS hand posture datasets II (NUS-II). The MIFD-Net significantly decreases its FLOPs and the number of parameters and reduces the complexity of the model while maintaining a high recognition rate compared with other gesture recognition models. The MIFD-Net can obtain high accuracy and strong robustness in different environments, lighting, and angles.

At present, road crack has become the main threatening factor affecting highway quality. The traditional manual detection method has low efficiency and produces larger errors. Aiming at this problem, this paper presents an improved object detection algorithm based on YOLOv5s network, fused with SE attention mechanism, which strengthens the important characteristic of the fractures of the target and suppresses general characteristics. Finally, we use the accuracy and recall rate as the evaluated parameters. Compared with the original network, the result has improved significantly, which greatly reduce the probability of crack leak fault detection. The location and type of cracks are marked out in the test results of this model, which effectively replaces the traditional manual detection method and optimizes the efficiency of road crack identification. After optimization, the lightweight network can be deployed on various mobile terminal platforms, making full use of the platform computing power, which owns high speed of identification and high precision, and has broad application prospects.

Localization and obstacle avoidance are important problems for indoor robots. Visual-based localization (VBL) is a promising self-localization approach that identifies a device's location in a 3D space by using cameras to see the device's surrounding scenes and objects. In this paper, we present a pictorial planar surface based 3D object localization framework. However, the image shaking on moving robot leads to localization accuracy reducing. In order to improve the localization accuracy on moving robot, the depth information from RGBD camera is involved to correct the pose calculation. Furthermore, in order to produce a more acceptable decision on obstacle avoidance, we also design an optimal path planning using RGBD camera based object detection. We have built an autonomous moving robot that can self-localize using its on-board camera and the PDPose (Picture Depth Pose) technology. The experiment study shows that our localization methods are practical, have a very good accuracy, and can be used for real time robot navigation. Moreover, compared with the traditional obstacle method, the optimal obstacle method produces better path planting result.

In modern military operations, in which electronic warfare plays an increasingly important role, radar jamming and anti-jamming technology is always an important research topic. With the rapid development of digital radio frequency memory technology (DRFM) in recent years, the development of deceptive jamming has made great progress, and numerous jams with good deception effect have been put into the actual battle and play an important role. Most of the current research on the classification of active deceptive jamming has chosen fewer types of jamming and is less generalizable. This paper addresses this problem by selecting a variety of types of spoofing jamming and compound jamming, and after extracting suitable feature parameters, using support vector machines and BP neural networks for classification, the trained classifiers have better performance in terms of accuracy and robustness, as well as generalizability, providing a theoretical basis for engineering applications.

According to the IPCC's Sixth Assessment Report (AR6) Working Group I report, Climate Change 2021, many changes in the climate system are directly linked to increased global warming, including extreme heat events, ocean heat waves, and increased frequency and intensity of heavy precipitation. The report, Climate Change 2021: The Natural Science Basis, released by the United Nations Intergovernmental Panel on Climate Change (IPCC) on Aug. 9. Climate change will intensify in all regions of the globe in the coming decades, with more frequent extreme heat and heavy rainfall events. The dual intensification of the intensity and frequency of extreme weather will have a dramatic impact on society as a whole and on humanity. Protecting the environment and slowing down global warming, thereby reducing the frequency and intensity of extreme weather events, will require concerted efforts by all in society. The main objective of this study is to understand the Internet-wide perception of extreme weather. This paper focuses on using a natural language processing (NLP) method, the Latent Dirichlet Allocation model, which can summarize and extract key information well. Our results found that comments on Twitter can be divided into 15 main topics.

The development of computer technology has far-reaching implications for the transmission of the intangible cultural heritage boat fist. In the general environment of regular epidemic management, E-learning mode of teaching is becoming increasingly popular. Distinguished from ordinary video follow-through practice, it provides real-time movement accuracy feedback while not limiting the user's learning venue. Therefore, in this paper, with the background of boat fist teaching, cultural dissemination and interaction design, based on Unity3D software and assisted by human body key point recognition technology, a boat fist teaching system is developed which is conducive to boat fist dissemination, cultural transmission, reducing learning costs and improving learning efficiency. The purpose is to spread and promote boat fist, a municipal-level intangible cultural heritage listed item in Qingpu District, Shanghai, so that more people can learn and understand them.

This paper uses CiteSpace to handle the anti-monopoly literature index data got from Web of Science, and the results show that the United States contributed most to the study of anti-monopoly; NYU, US Dept Justice and others are important research institutions in this field; LAW J, J POLIT ECON, TEX LAW REV and others are the most important journals; Easterbrook FH, Posner RA, Bork RH and others are the most important authors.

The traditional medical voice question and answer interactive system can not meet the core needs of patient consultation service, and the patient's satisfaction with the system answer is low, so the research of medical voice question and answer interactive system based on speech recognition technology is proposed. In the hardware of the system, the control core of the system is the MCU module, which controls the normal operation of the whole system, realizes the conversion of the analog signal and digital signal through the audio input device, designs the serial communication device to transmit the data information to the computer, and uses the speech recognition technology to realize the software part of the system. And through the system test, the medical voice question and answer interactive system based on speech recognition technology is compared with the traditional medical question and answer interactive system. The experimental results show that the average response time of the medical voice question and answer interaction system based on speech recognition technology to the user request is far less than the traditional system, there is no number of error requests, and the performance is better than the traditional system. Therefore, it can be proved that the performance of the medical question and answer interactive system based on speech recognition technology has reached the expected standard, and can meet the basic needs of users for medical consultation services

In recent years, with the development of machine learning as well as deep learning, named entity recognition has also been rapidly developed. Currently, with the huge saturation of the talent market, resulting in a dramatic increase in the number of resumes, it has become more difficult to screen resumes, and company personnel have taken on a huge burden of screening work. The application of named entity recognition on resume parsing can greatly shorten the time of resume screening and apply more time to the screening of target talents. Based on the problems faced by named entity recognition in resume parsing, this paper proposes traditional machine learning models HMM, CRF and deep learning model BiLSTM+CRF to solve them. After a series of experiments, the effectiveness of the proposed method is proved.

Aiming at the problem that the advanced arresting device is difficult to obtain the landing speed of the aircraft in real time, this paper predicts the speed of the aircraft through the self-encoding fuzzy inference system (AE-ANFIS). Firstly, the working principle of the advanced arresting system is expounded, and the sensors directly related to the aircraft speed measurement are analyzed. And filter auxiliary variables through feature extraction and maximum information coefficient (MIC); then predict acceleration through adaptive fuzzy neural network (ANFIS); finally, for the problem of over-fitting caused by the large number of ANFIS rules, an auto-encoder (AE) method is proposed. Data dimensionality reduction is performed by extracting abstract features, which effectively improves the prediction accuracy of ANFIS. The experimental results show that the method proposed in this paper can fit the aircraft speed well, and the accuracy is better than traditional ANFIS and BP, LSTM, GoogleNet, AlexNet and other algorithms.

Optimization of the terminal distribution route of fresh products is very important to reduce the operating costs of fresh e-commerce enterprises. In this paper, we take Freshhema as an example to analyze the terminal distribution mode and find out the problems existing in the distribution process. Through conditional hypothesis, parameter setting and adjustment, the path optimization model of terminal distribution problem is constructed, and the algorithm is designed to solve it. Finally, in view of the practical problems of Freshhema, a reasonable terminal distribution scheme of Freshhema is put forward. Meanwhile, it also provides a reasonable idea for the terminal distribution route optimization of fresh e-commerce under the new retail background.

Sentence representation is a typical problem in NLP, which is to use a fine vector to encode the sentence, so that the sentence can contain copious semantic. A high-quality sentence representation benefits a wide range of NLP tasks. Although BERT-based pretrained model performs well on many downstream tasks, it’s sentence representation acts poorly on semantic textual similarity (STS) task. In this article, we propose ConBG, a Contrastive Learning Method for Chinese Sentence Representation Based on Bert and GCN, which is to encode the sentence by a model combined with Bert and Graph Convolutional Network which is to incorporate syntactic information. Then we use data augmentation strategies to create samples, and adopt contrastive learning technique to train the model in a unsupervised way. Experiments on Chinese STS datasets demonstrate that ConBG exceeds previous work of over 1% on average.

The online education platform is a typical Internet scenario. As an online education platform, the platform faces many users, including students, teachers, and adult users. The system positioning determines that the platform needs to deal with high concurrency issues and consider the scalability of the system. When the number of users is too large, a single development framework cannot solve the problem well. Therefore, considering this type of typical Internet scenario system and taking system load dynamic planning as the research content, the MTBBO algorithm is proposed in the paper to deal with this problem. Through the comparison between MTBBO and single-task single-objective models GA, DE and BBO, the superiority of the algorithm in this paper is verified.

Classroom behavior is an important criterion for evaluating instructional efficacy. In comparison to other behaviors, the challenge of classroom behavior detection is primarily influenced by ambient light variables and the presence of too many targets to recognize, resulting in missed detection. Recent research has demonstrated that information about the human skeleton can be used to identify classroom conduct. As a result, we present an enhanced yolov5-based skeletal recognition system for detecting classroom behavior in this paper. First, the YOLOv5 detection algorithm is improved to extract target prospects for the problem of missed detection; then, the human skeleton information is obtained using the Alphapose framework; finally, the skeletal data is sent into a two-stream adaptive graph convolution network to allow for the accurate recognition of various classroom behaviors. According to extensive tests, the detection algorithm based on bone recognition improves detection accuracy and lowers the false detection rate.

With the rapid development of computer technology and communication technology, computers have been applied to all areas of people's lives. In the field of education, it is a necessary function for the intellectualization of the education system to realize the automatic marking of examination papers by computer. In intelligent marking, the traditional method is used to segment the subjective question area of the test paper, which can not effectively segment different types of answer sheets and has the defects of low segmentation accuracy. The edge detection method is used to correct the subjective question area of the answer sheet, and the boundary box of the subjective question part is accurately screened by positioning the top, left, and bottom positioning lines of the answer sheet to complete the segmentation of the subjective question area in marking. The experimental results show that the correct rate of the new method is more than 95%, which is much higher than that of the traditional method. The new method can be widely used in intelligent marking work, and can effectively improve the efficiency of intelligent marking.

8-DOF high-precision welding robot innovatively adopts the flexible integration technology of robot and peripheral equipment to provide systematic intelligent welding solutions for enterprises in need. The welding robot is flexibly integrated into the welding workstation with positioner, welding machine system, wire feeding unit, tooling fixture, gun cleaning station and other peripheral equipment. According to the actual production needs of the demanding enterprise, the welding robot completes the path planning, the robot controller system coordinates the control system to collect sensor signal instructions, communicate with the robot, control the robot and peripheral equipment to coordinate actions, and ensure the efficiency of the whole system complete welding work with high quality.

The 2019-nCoV can be transmitted through respiratory droplets and other methods, which greatly endangers public health security. Wearing masks correctly has been proven to be one of the effective means to prevent virus infection, but limited by the complexity of practical application scenarios, the wearing of masks still relies heavily on manual supervision. Therefore, a fast and accurate face mask wearing detection method is urgently needed. In this paper, a mask detection algorithm based on improved YOLO-v4 is proposed as a solution to the problems of low accuracy, poor real-time performance, and poor robustness caused by complicated environments. In addition, a number of different training approaches, such as mosaic data augment, CIOU, label smoothing, cosine annealing, etc., are introduced. These techniques help to increase the training speed of the model as well as the accuracy of its detection. With a fast-training model, the model will be able to detect and compare the results of samples from different scenarios. The experiment will compare front and side faces, different colored masks, scenes of varying complexity and other perspectives in a systematic way. The experiment's result was able to reach 99.38 % accuracy after the model was trained using data from a variety of face masks being worn. Experiment results, both quantitative and qualitative, indicate that the method can be adapted to most scenarios and offers effective ideas for improvement.

In recent years, deep learning technology has achieved remarkable results in stock price prediction. Improving the performance of prediction models by enhancing the data characteristics with news data is a hot research direction. This paper proposes a possible improvement in stock prediction based on financial news by analyzing the effectiveness of news. Firstly, this paper only trains a Long Short Time Memory model with stock price data and gets a baseline result. After that, a series of experiments are taken with two companies' stock price data: Microsoft and Apple. In those experiments, this paper explores the correlation between stock price and emotional news scores and the news's time delay effect. By comparing the impact of whether to introduce emotional scores on the prediction results of the model and the differences in the prediction results of the model under different time delays, this paper obtained some conclusions. The experiment results show that the news time effect differs in various companies. In addition, some existing correlation measurement methods cannot become the basis for measuring whether the data to be supplemented (emotional scores of news) is helpful. Based on this, news data enhancement is not always practical.

With more and more applications of deep learning, including neural networks, the lack of explanatory make it easy to under external attacks. This paper mainly focuses on adversarial attack, that is, by adding slight perturbation to the input data, which cannot be detected by human being, can lead to a wrong output of the model and maximize the model’s prediction error, resulting in a distrastic decline in the performance of the model, including prediction accuracy, etc. But so far, there is still no sufficient theoretical support for why adversarial attack that can not be detected can lead to serious performance degradation of neural network models, some attempts at explaining this phenomenon focused on the reason of over-fitteing or the linear or unlinear nature of the neural network. In this paper, several experiments based on Fast Gradient Sign Attack Algorithm to resist adversarial attack are designed and implemented towards the model of neural network on the tasks of the image recognition and classification, and some regular experimental results are obtained. The experimental results provide some evidence for the argument that the reason why adversarial examples and adversarial attack can work is over-fitting, but it is possible that the evaluation methods can not measure such special kind of overfitting.

Livestock modernization is the front-runner and wind vane of agricultural modernization. Smart animal husbandry for living individuals such as sheep has been widely researched and practiced as a typical application solution of smart agriculture. How to accurately identify different individuals in the same kind of target in the shortest possible time is one of the key problems facing the field of visual recognition and matching at present. In this study, a spatial invariance method is proposed to match multi-target individuals. In the paper, the set of line segments is taken as the unique feature of the target, and a threshold is proposed as the criterion to judge the difference between targets. Experiments show that the accuracy of line segment measurement can reach 99.00% on average; the recognition accuracy reaches 100% when the difference between target individuals is large. The method in this paper has some research significance for the identification of clustered cattle and sheep individuals in smart farming and animal husbandry.

Segment-powered linear motors (SLM) have high reliability, unbalanced operating data samples and few abnormal samples, making it difficult to detect anomalies through supervised classification models. Therefore, we propose an anomaly detection method based on the reconstruction model, which uses the difference between small reconstruction error of normal samples and large reconstruction error of abnormal samples to detect anomalies. Meanwhile, in order to make the reconstruction model better reconstruct the input samples, this paper introduces adversarial regularization to improve the training effect. The proposed model uses normal samples for training, and unknown state samples for testing. The anomaly score is obtained through the reconstruction error between the input sample and the reconstructed sample, so as to judge whether the input sample is abnormal. Experiments are carried out on the actual operation data of the SLM, which proves the effectiveness of the proposed method in the anomaly detection task of the SLM.

With the rapid development of ESG investment, more and more world-renowned institutions have started to establish ESG rating systems to measure the sustainable development of listed enterprises, but due to the late start of ESG in China, and the current domestic research on ESG, rating scores is limited to the impact of ESG rating scores on corporate financial risks, corporate financing constraints and the reasonableness of ESG rating systems. Therefore, this study establishes a localized ESG rating system in China based on China's policies and national conditions, which helps China's ESG development to advance rapidly and also provides an effective reference solution for the Chinese government to monitor the sustainable development behavior and development of listed enterprises.

Straw burning is a harmful and dangerous method to treat the residue of crop straw. At present, remote sensing index is widely used in the detection of incineration area. This paper tries to discuss the factors and specific ways that affect the discriminant effect of combustion index. In this paper, we used Landsat 5 image data taken in 2011 for the study area of Henan Province, combined with MODIS flash point data from May to June 2011, and then analysed the three factors on the impact of straw burning index white (burning area index). After dividing different influencing factors, the differentiation index of BAI was calculated to judge the effects of influencing factors. The results show that for the recognition indexes of different bands, the value of the near infrared band is the largest. At different times, the discrimination index decreased significantly around the 10th day. In the planting system, the effect of distinguishing two crops a year is better. It is concluded that BAI index differentiation effect is the best in the near infrared band in 10 days before the combustion of one year and two seasons.

In the traditional exhibition, it is limited by the space of the venue and the scarcity of precious objects. With the help of 3D scanning technology and mixed reality technology, various restrictions can be reduced to see more exhibits. Use 3D scanning to obtain 3D data of items, generate 3D high-precision digital models, and then display interactively with exhibits through mixed reality technology.

The rodent infestation problem is currently one of the important factors in the degradation of grassland in the Sanjiangyuan area. We need to infer the degradation of grassland by the area of grassland being gnawed, and thus provide help for grassland restoration work. To this end we have designed a DeeplabV3+ based mouse infestation scene segmentation method. On the basis of Deeplabv3+, different backbone feature extraction networks are adopted, and attention mechanism is introduced into the backbone to improve the accuracy of feature extraction and solve the problem of sample imbalance in our self-made dataset. For the training and validation of this network, we used a self-developed photographed and produced dataset of the distribution of mouse holes in the grassland pastures of Haibei, Qinghai Province, which contains various features of plateau mouse infestation. The model improvement resulted in a significant reduction in the training time of Deeplabv3+ on this dataset, and a certain degree of improvement in segmentation accuracy.

Aiming at the problems that the previous entity relation extraction model has insufficient dependencies between words, the recognition effect of overlapping entities is low, and the triples caused by single decoding force an unnecessary order, this paper proposes a deep learning-based method. The method uses a pre-trained model to extract sentence features, uses the Span method with stronger ability to extract overlapping entities for entity extraction, and uses a deep multi-fork decoding tree to implement parallel decoding. The experimental results on the CoNLL04 and ADE datasets show that compared with other relation extraction models, the F1 value of the model in this paper has a better improvement, and it also verifies the effectiveness and generalization ability of the model in this paper.

With the continuous exploration of unmanned technology in recent years, techniques for road vanishing point detection have emerged, mainly including the exhaustive searching based methods, RANSAC based methods and expectation maximization (EM) based methods. This paper propose an exhaustive searching and RANSAC based algorithm called vanishing point detection which is capable of detecting vanishing point in road images. Our proposed algorithm first uses the LSD algorithm to determine the vanishing point in the road and detects the line segments in the filtered image; then we detect the candidate vanishing points (referred to as VPs) and determine the final VPs by mapping the candidate VPs to the sphere. Compared with other algorithms on more than 120 test images, the error of our method is within 5%. The effectiveness of our approach is demonstrated by quantitative and qualitative experimental results. And our approach offers a new and effective algorithm to image recognition of driverless technology.

CAVs, or connected and autonomous vehicles, have the potential to revolutionize the transportation system by making vehicle driving safer, more comfortable, and more efficient. As a result, there is the question of ensuring that CAVs and conventional vehicles can coexist insightfully. Our goal in this research is to understand better how CAVs can coexist with non-autonomous vehicles, i.e., human-driven vehicles (HVs) on city streets. To use as little energy as possible while maintaining high levels of safety, we apply the optimal control strategy and study the problems when vehicles in mixed traffic flow attempt to merge at on-ramp areas. To verify this approach, we conduct simulations with Python and SUMO. The effect of CAVs on fuel consumption is investigated under different penetration rates and traffic flow volumes. Energy efficiency gains become more significant as CAV penetration increases, whereas traffic volume decreases its significance.

Today, with the proliferation of electronic system terminals and the ever-changing needs of people, a variety of reminder services have emerged, but existing reminder services are still mostly limited to smartphones, and smart wearable devices (such as smart watches with functions such as heart rate monitoring and alarms), and the relevant industries and markets are still expanding. This paper introduces a customized integrated smart carrier reminder platform, which, given the limitations of current carrier reminder functions, shifts the focus of innovation from the commonplace smartphones and wearable devices to not yet widely noticed carrier reminder services, and proposes an integrated, multifunctional smart reminder system; The idea is to receive information from the user through a variety of input methods (contact input, contactless input, programmed input, etc.), collect and process the data through a central processing system (internal and external sensors) and finally output the feedback (e.g. light, sound, smell, touch) from the output system. The advantage of this study over existing carrier alert services lies in its platform-based operation mode and high level of user interactivity. A use case is demonstrated to verify the usefulness of the proposed system.

In the case of the existing power grid partition, the actual transmission between each node has fluctuations, which will lead to low electrical modularity of each partition. To solve this problem, this paper proposes a method of reactive power and voltage partitioning based on fuzzy clustering. The transmission state between any two nodes of the power grid is analyzed in detail by using the index parameter electric coupling strength which directly reflects the electric characteristics of the power grid, and then the fuzzy clustering method is used to realize the partition processing of the reactive power and voltage of the power network according to the analysis results. According to the relationship between the power flow coefficient and the node electric coupling strength coefficient obtained by the Newton-Raphson iteration method, the buses corresponding to the coupling strength coefficient are clustered based on the integrated grid load. Finally, the reactive voltage partition is realized according to the scale standard of 0. 2. The test results show that the electrical modularity of the partition results of the design method is always stable above 0. 60, which is better than the comparison method and has high feasibility.

As a chronic disease endangering human life, AIDS seriously hurts human health and living. The study on the AIDS prediction model can provide a critical basis for the prevention and control decision-making and resource allocation. This paper selected the monthly AIDS incidence cases and mortality cases for ten years (July 2012 - June 2022) as the research objects. We analysed the data, combined the national AIDS transmission characteristics, and established the prophet model. Realized the Prophet Model modelling, verification, and prediction of monthly AIDS incidence cases and mortality cases and forecast them for the next six months in China. The actual data showed a completely fluctuating trend. In addition, from the forecast results, this trend is still relatively explicit in the future. The MAE of incidence cases is 400.68, which accounts for 7.6% of the mean actual values. The MAE of mortality cases is 74.22, which accounts for 4.47% of the mean actual values. Effective modelling and prediction of HIV incidence cases and mortality cases enable AIDS patients to recognize the harm in a time and effectively prevent and control its occurrence.

Plentiful information is contained in the massive gene expression data obtained by DNA microarray technology. Efficient computer analysis methods are beneficial for extracting helpful information quickly and accurately and assist in understanding biological phenomena from enormous data. In this paper, through the integration of fisher-score in filter layer, recursive feature elimination and logistic regression in embedded layer, FSRL is proposed to search latent biomarkers in various cancers with high mortality rates. Compared with the currently popular online analysis tool GEO2R, FSRL has higher classification accuracy. FSRL performs better in evaluation index than five methods and dramatically enhances calculative efficiency. In prostate datasets, evaluation indicator such as accuracy of FSRL, is 40.1% higher than the average. Since biomarkers obtained through multiple cancer sets are more reliable and repeatable than single analysis, cluster analysis is conducted on six cancers with high mortality rates, and 13 genes are screened to form potential genomic biomarker modules. The genes are validated through literature review, GO analysis, and functional pathway retrieval to provide information on carcinogenic mechanisms. They can be used as a decision support system for potential biomarkers to help to narrow scope of biotic experiment scope and detect multiple cancers in targeted anti-cancer therapies.

Taking evolutionary game theory as the main research tool, on the basis of describing the relationship among listed companies, accounting firms and regulatory departments, this paper analyzes the financial fraud behavior of listed companies by constructing the game model of "listed companies, regulatory departments and accounting firms", and finally uses MATLAB to simulate the game model. The research shows that the financial fraud of listed companies is not only driven by their own interests, but also affected by multiple factors such as the punishment of the regulatory authorities, the regulatory success rate, and the willingness of accounting firms to cooperate. On this basis, from the three perspectives of listed companies, accounting firms and regulatory departments, this paper puts forward countermeasures to curb financial fraud of listed companies.

Some recently proposed data augmentation methods have been used to solve the overfitting problem of neural networks and are gradually becoming a research focus in deep learning. These data augmentation methods have been widely used in tasks such as image recognition, target detection, and image segmentation. However, for the overfitting problem of neural networks on video data, the existing data augmentation methods have the limitation of feature dimensionality; they can only affect the spatial features of the training samples. Moreover, the current methods lack the filtering effect on the temporal information of training samples; the training samples after data augmentation still have more rea information, and the video behavior recognition model still suffers from the overfitting problem. Therefore, this paper proposes the ShiftMask data enhancement algorithm. The method in this paper uses a new masking approach to correlatively mask the temporal and spatial dimensions of video data to help the model identify the subject of the action and reduce the subject and scene learning bias of the model during the training process. Moreover, this paper utilizes a grid-like mask of varying sizes to preserve the data's fine-grained features. Finally, experiments on the HDMB51 and TobaccoFactory datasets improved the recognition accuracy of the I3D model by 4.5% and 2.9%, which were better than other mainstream data enhancement methods.

The appearance of big data provides many changes to our daily lives. Among these changes, the innovation and application of big data in higher education are greatly remarkable. Therefore, the paper firstly analyzes the background of big data appearance. Then the paper describes the main application in higher education, which includes improving college planning, promoting college development, obtaining the teaching reality, enhancing teaching efficiency, optimizing learning experience, improving learning quality and boosting science research. After that, the paper provides two technologies of big data analysis: educational data mining and learning analysis. The applications of the two technologies in higher education mainly include classification, forecast, clustering, outlier detection and others. At the same while, the paper also describes the challenges of big data application in higher education, which are epistemology, methodology, technology, law, ethic and other dilemma.

Topic modelling approach is widely used for text data mining in NLP(Natural Language Processing). Text mining has been used for analysis of ICH (intangible cultural heritage), where Cantonese opera is a representative ICH of Lingnan culture. This study retrieved news content on Cantonese Opera and used machine learning analysis (LDA topic modelling) method to find out the distribution of the topics. Four main themes are concluded: the development, cooperation, and inheritance of Cantonese opera(taken up to 45.1% in all data); The traditional form(23.5%); Innovative forms(18.3%); Education and cultural inheritance of Cantonese opera(13.2%). This research further explored how to better promote Cantonese opera by analysing the topics as well as the data, and suggested that emphasis should be placed on the innovation of traditional elements in Cantonese opera, keeping them close to life, and education.

In the process of the construction of China's new power system, we will vigorously promote the research and development of UHV power equipment and the wide application of power electronic devices. UHV power equipment has complex insulation structure and huge volume, bear impact energy load of wind power and photovoltaic of new power system for a long time. It will cost a lot to carry out on-site operation and maintenance tests. Digital twin technology is becoming more and more perfect, and new power system construction is gradually introduced from automobile, aviation and other manufacturing industries. Based on this, this paper introduces the digital twin technology into the high-end power equipment of the new power system, and carries out on-site operation and maintenance simulation test and functional response analysis under high current, high voltage and multi harmonic loads according to its twin model. From the four sensing dimensions of mechanical vibration, gas composition, optical vision and electrical parameters, the improvement of intelligent sensing technology of new power system equipment is analyzed, and the interaction between on-site operating parameters and digital twin model data is realized. On the other hand, GPU computing power expansion technology supporting digital twin multi-source sensing technology is proposed, which can effectively support the dynamic behavior simulation monitoring of equipment from 10-5 seconds to 103 seconds, and the operation life evaluation strategy of high-end equipment is proposed. This paper focuses on the 3D construction of the digital twin model of the high-end equipment of the new power system, and its research method can be extended to the construction of the whole network digital twin model of the new power system. The research results can provide theoretical guidance and technical reference for the application of digital twin technology in high-end power equipment scenarios, and effectively support the safe and stable operation of the new power system with "double high characteristics".

This paper first analyzes the data visualization teaching in the university classroom, then draws the conclusion that the Matplotlib in Python language is suitable for data visualization teaching and data visualization teaching will be the inevitable trend of the development of information technology. Finally, an example is given to discuss the teaching practice of Matplotlib in data visualization teaching.

Satellite remote sensing images have the problems of large image scale, dense arrangement of segmentation targets and different directions. And the specifications and clarity are far from the natural images, resulting in difficulty in feature extraction. Therefore, the detection accuracy of Mask-RCNN is poor when applied to remote sensing image instance segmentation. In this regard, an improved Mask-RCNN algorithm is proposed. First, a deformable convolution kernel is introduced into the back bone network to adaptively change the theoretical receptive field. On this basis, the FPN module is modified, and feature layered fusion is introduced to further improve the feature extraction capability of the model. At the same time, the Soft-NMS algorithm is used to screen the target candidate frame. Validated using the iSAID dataset. The experimental results based on the data set show that the improved algorithm has higher detection accuracy than the original algorithm, and the missed detection rate is reduced.

Due to the particularity of tobacco as a commodity, the problem of cigarette trans-regional outflow has always been a difficult problem in the standardized management of the tobacco market supervision, mainly manifested in the lack of detection means, low identification accuracy, lagging early warning results and so on. To solve the above problems, this paper proposes a detection model TMSS (Tobacco Market Scientific Supervision), which can accurately identify the risk points related to tobacco operating by constructing an isolation forest algorithm, combining principal component analysis and grid search techniques, using multisource data fusion. This paper focuses on the working principle and algorithm design of the model TMSS. The accuracy and reliability of the TMSS are verified by selecting relevant application scenarios in Shiyan city, Hubei province. Through experimental verification, TMSS can effectively identify the risk points cause the transregional outflow phenomenon in the tobacco operation process in advance, which provides a significant breakthrough method and tool in the field of scientific supervision of tobacco market.

Unstructured data e.g. images and videos are widely used in the medical field. Because the generative adversarial network (GAN) has the ability to process images with fewer labels and better feature extraction, the application of GAN in surgical video can promote the development of medical fields such as surgeon training and telemedicine. From three aspects, surgical procedure, video enhancement and imitation learning, the article summarizes the current applications of generative adversarial network (GAN) in surgical video processing. The first is two specific applications in terms of surgical procedure, step prediction (i.e. Supr-GAN) and surgical image generation. Second is about video enhancement. Based on the real-time performance and video processing effect, the paper introduces three types of applications in real-time video, respectively network delay, sharpness improvement and device recognition, and two processing methods to non-real-time video with the mirror reflection problem or the smoke problem. Finally, the paper also summarizes two applications of generative adversarial imitation learning (GAIL) in surgical videos, which focus on surgical suture (i.e. Motion2Vec) or selective catheterization simulation. The summary indicate that GAN currently focuses on minimally invasive surgery video processing, which means the type of surgery is relatively monotony. In addition, GAIL is seldom used to simulate learning based on surgery videos. Therefore, GAN, especially GAIL, still has a broad prospect in the application of surgical videos.

The digital transformation of education reshapes the daily practice and meaning generation in the field of education and teaching, promotes teachers' teaching decision-making out of the "experience-based" value judgment mode, and combines the value advantages of "Data driven", so as to make the decision-making more sufficient, refined and efficient. The value of data based decision making1 (DBDM) in the digital age is more to help the continuous improvement of teaching and learning, and help to promote the realization of high-quality teaching objectives in the new era. To promote the implementation of DBDM, we need to improve teachers' data literacy, strengthen the research of data based decision making mode, promote the application of multi-source heterogeneous data fusion based on scenario requirements, and create a culture of data use.

Nowadays, with the rapid development of entertainment industry, filter has almost become an indispensable function in photography. However, most of the filters today are not directly change the style of an image, and even those techniques that can achieve image style transfer are difficult to be applied by users of other disciplines because of the complexity of their algorithms. In this study, a Generative Adversarial Network named CycleGAN, is used to transfer normal landscape images to images with the style of Van Gogh’s paintings. By using this model, two generators and discriminators will be formed. For the generator, they both generate a domain of images to the other domain, transfer them back to the original domain, and to make sure the generated images will not be differed so much with the original images. In this way, the generator can learn the important features of a specific style of images and apply them to another set of images with a different style. Experimental results show that the model does well in transferring normal landscape images to images with the style of Van Gogh’s paintings, but for the transformation from the style of Van Gogh’s paintings to landscape images, the effect of the model is not very ideal.

Under high-tech conditions, the communication coverage is wide, the amount of information transmission is large, the structure is complex, the timeliness requirements are high, and the accuracy guarantee is difficult. In view of the new problems faced by the communication network operation and maintenance, relying on the network observability system and massive operation and maintenance integration of data resources, explore the data-driven multi-agent decision-making technology, and improve the whole process communication planning guarantee, event analysis and disposal, and prediction and decision-making ability, obtain and maintain information superiority under the fierce information confrontation.

Shore based information service is the key technology of ship shore cooperation theory for intelligent navigation of ships, and is an important guarantee means to support autonomous navigation. This paper proposals to apply maritime service technology to navigation, relaying on modern digital communication technology to establish a real-time dynamic, location-based, on-demand service geographic information service operation system. Tests show that shore-based information service technology can meet the demand for dynamic and maritime safety information in modern navigation, and has a position Effect on improving crew efficiency and improving navigation safety.

Knowledge graph is widely used in the artificial intelligence field such as intelligent search, intelligent recommendation and intelligent question answering. Financial knowledge graph (FKG) can achieve the purpose of intelligent cognition of investment in financial market, to realize efficient mining and intelligent application of financial data. In this paper, FKG is applied in stock investment analysis and intelligent recommendation, and the design idea of multi-dimensional construction of stock finance knowledge graph (SFKG) is proposed in combination with investment behavior. We conduct research on intelligent recommendation algorithms via knowledge abstraction and generalization. A-share market data of China is used for testing and verification, and an intelligent visual system with auxiliary decision-making function is implemented finally.

With the rise of artificial intelligence and other technologies, unmanned driving has become an important direction of future automobile development. The ground extraction of roads is a key task. The traditional method of ground extraction is limited to flat roads and the speed is slow. In view of the under-segmentation problem between the ground point cloud and the multi-objective object point cloud, we propose a ground segmentation algorithm for point cloud data based on multi-region segmentation. In this paper, the ground data can be accurately found by using the method of concentric region division for point cloud data. At the same time, it also has good robustness for uneven road surface, so as to provide a good foundation for the subsequent segmentation of road obstacles. For the extraction of road obstacles, we adopt the DBSCAN clustering method based on sub-region. According to the characteristics of near-density and distant point cloud, we adopt adaptive parameter selection for different regions to improve the accuracy of obstacle extraction. In order to verify the results, the use of KITTI data set on ubuntu18.04 using ROS system to complete the test, can maintain about 28fps processing speed.

Computed tomography (CT) provides a three-dimensional view of the patient’s internal organs. X-ray imaging offers a two-dimensional view for patients. X-ray images are more commonly available and less costly than CT, and the radiation dose to the patient is significantly reduced. Traditional CT imaging methods require projection with hundreds of X-rays for a full body scan. An end-to-end generative adversarial networks (GAN) network approach, i.e., TPG-rayGAN, was proposed for reconstructing lung CT volumes directly from biplane X-ray images. In this work, CT was reconstructed with ultra-low radiation. Densely connected networks and Transformer networks were connected in parallel to extract features. In addition, the perceptual loss function was added in the loss function section. The experimental results show that high-quality CT can be reconstructed from X-ray images using TPG-rayGAN.

In the era of big data, digital economy arises at the historic moment, and strengthening consumption governance is its important proposition. Massive data provides rich resources and reliable support for consumption governance. Driven by big data, consumers' consumption behavior is more rational, and their consumption choices are more vulnerable. At the same time, consumers' dependence on brands gradually decreases, but personalization gradually increases. The change of consumption trend also brings great opportunities for consumption governance. Information disclosure makes consumption governance more intelligent. Improving services to make consumption governance more precise. Improving quality makes consumption governance more rational. Customer first makes consumption governance more humanized. Therefore, this paper explores the consumption governance path by using the data mining (DM) method through sampling survey. The results show that we need to continuously improve the efficiency of consumption governance by strengthening the technological innovation, improving the safety factor, strengthening the regulatory regulations, and improving the coordination of policies.

Although voids play an important role in Chinese paintings, when generating Chinese ink wash paintings from real life photos of horses, background detection and void-leaving remain challenges in producing more credible paintings. To address the problem, a model with a two-stage framework is proposed in this paper. The framework divides the generation process into background lightening and style transformation. In the first stage, by training a pix2pix model with paired original photos and background-lightened photos, the pix2pix model is enabled to detect and lighten the background correctly in the style-transfer process, hinting where voids should be left. In the second stage, a Cycle-GAN model trained with unmatched photos and Chinese paintings achieves the style-transfer from pre-processed photos to Chinese ink wash paintings. In preparation for training the stage-Ⅰ model, the photos from a given dataset is processed in batches to get corresponding background-lightened photos. During the training of the stage-Ⅱ model, original photos instead of processed photos is used as comparatively data augmentation, making the model more robust and independent. Compared to the baseline model, the proposed model reaches a higher accuracy in both void-leaving and detail-enhancing in the experiments, resulting in more credible and delicate generated Chinese paintings.

With the rapid information and communications technology growth and continuous invention, the concept of parallel computing has become the core of computer science, and its capabilities are continually documented in providing high processing solutions. The graphic processing unit (GPU) plays a crucial role in deep learning and parallel computing. The primary purpose of this paper is to discuss two concepts of deep computing: data parallelism and model parallelism. Additionally, the discussion will delve into their importance and associated challenges in parallel processing. Furthermore the paper summarises the Residual Network (ResNet) system with the basic principles and formulas. Establishing efficient parallel algorithms for GPU processing is necessary. This paper evaluates the performance and speed between single-core and multi-core working environments for two kinds of convolutional neural network structures with different types of GPUs. The first one is LeNet, and the other one is VGG16. There are two final results of this experiment. By analyzing the data of performance, VGG16 has better performance than LeNet. By analyzing the data of speed, the speed of LeNet is faster than that of VGG16.

As the open source ETL tool with the largest number of users, Kettle's powerful and concise functions are welcomed by the majority of ETL practitioners. ETL is the process of extracting, converting and loading data from the data source to the target. Based on the enterprise data model, ETL builds a reasonable data storage mode, establishes an enterprise data exchange platform, meets the data exchange needs between various application systems, provides all-round data services, and meets the data support needs of enterprise decision-making [1]. In addition to formulating the corresponding extraction, transformation and loading tasks, ETL tools are also particularly important for the formulation of scheduling monitoring. For industrial economic data, benefit data and production data are generated periodically (such as daily, weekly and daily). Through the scheduler, users can more easily realize ETL task scheduling management, task execution performance and other data monitoring and analysis. However, Kettle's scheduling and monitoring function is very weak. This paper will use the scheduling system to monitor and schedule tasks to improve the efficiency of data integration.

CAPTCHA (Completely Automated Public Truing test to tell Computers and Humans Apart) is a technology used to protect network security, distinguishing humans and machines by tests that humans can easily pass but machines fail. However, with the wide application of CAPTCHAs, many cracking methods have emerged. Specifically, Deep Learning (DL) have played an important role in in destroying CAPTCHA, which brings a great threat to the security of CAPTCHAs. Therefore, it is of great significance to improve the security of the CAPTCHAs against cracking methods based on deep learning technology. This paper focuses on image-based CAPTCHAs which is the most widely used type, and proposes an image-based CAPTCHA generation method based on adversarial examples. Starting from the actual application scenario of CAPTCHA, we improve it on the basis of I-FGSM (Iterative Fast Gradient Sign Method) by randomly transforming the input samples and performing secondary processing to adversarial noise. Experiments show that our method can effectively improve the robustness and transferability of the CAPTCHA.

The fundamental issue in network traffic measurement is data stream processing, and the main technique is estimating the frequency of various things to determine the top-k elephant flow. Finding top-k flows is crucial for the use of network management technologies including traffic engineering, DDoS attack detection, network anomaly detection, and congestion control. The expansion of high-speed networks has led to an increase in network traffic, making it important to create a quick and accurate online identification top-k method that uses less bandwidth. Our research revealed that it is challenging for current algorithms to simultaneously attain high precision and high throughput when memory is limited. To address this issue, we propose a brand-new data structure called ActiveHolder that combines counter and sketch technology and uses a replacement decay scheme to maintain the frequency values of the top-k flow items. This significantly reduces the impact of mouse flow on elephant flow. The experimental findings demonstrate that the ActiveHolder algorithm outperforms the state-of-the-art technology by achieving high precision and high processing rates with minimal memory cost.

The immersive visualization method expands the interaction space of traditional visualization, allows users to perceive information in multiple channels, and enriches the dimensionality of information. Embodied cognition provides theoretical support for the interaction process of immersive information visualization by emphasizing the role of the body and the environment. Based on the five elements of interaction behavior, the paper builds an interaction analysis framework of "Embodied-Behavioral" interaction framework (EB) from the perspective of embodied cognition, which includes three aspects: perception, interaction, and understanding; Introduces the current research in immersive visualization and analyzes the interaction process from different perspectives. Several aspects need to be considered when designing immersive visualization interactions, such as maintaining temporal and spatial consistency, promoting fault tolerance, and reasonably applying environmental information.

Due to the different parameters in the algorithm varying degrees of accuracy and complexity are affected, besides, as the number of categories in the data set increases, gradient explosions in classifiers cause unwanted crashes in early training. Therefore, synthesizing high-resolution photo-realistic images has been a long-standing challenge in machine learning. In this paper, a new idea is provided that the aim of providing more accurate image synthesis, and giving more powerful feedback (i.e. accuracy) Briefly, this study by adding a new generator layer and modifying the number of channels and the activation function to improving the performance of composite images. At the same time, by using the control variable method (i.e. channel number and activation function) to analyse respectively. Tring the situation when the number of channels attempted is 64.32.16 and the activation function is ReLU and LeakyReLU respectively, to accumulate different numbers. Finally, visualize different situations, and show them in the paper. Experimental results indicate that by adding a new generation layer and changing the parameters separately, the accuracy of the generated images is better than before. The result will be the best when the number of channels is 16, and the activation function is LeakyReLU.

In criminal and victim identification, when information such as fingerprint and facial images cannot be obtained, the development of other new bio-metric recognition has become an important task. It is found from investigations that the lower leg is often exposed in an evidence image. If we can use the information to identify the identity of the suspect, it will be a great break-through for criminal investigation. Unlike faces or fingerprints, the image of a lower leg does not have many discriminative features. Therefore, capturing more information from an image has become the key to the success of recognition. Despite the rapid development of deep learning in recent years and the wide application of convolutional neural networks in biometric recognition, they are very dependent on a large amount of training data. In this paper, we propose a recognition method based on the combination of chaotic mapping and local binary pattern (LBP), which captures the nonlinear features of the image. Experimental results show that the method proposed in this paper has a better recognition rate and an encouraging performance.

Due to the rapid product market cycle in the consumer goods industry, the development of digital technology has promoted the market layout of consumer goods. It is necessary to carry out brand publicity and strategy research by means of big data, the Internet of Things, 5G and other digital means, and form interactive communication with consumers. With the popularization and application of mobile Internet and mobile terminal equipment, great changes have taken place in people' s consumption behavior and information acquisition mode. Marketing technology and methods have changed from offline to online. Through the current situation analysis and data research, this paper puts forward new ideas and opinions on the digitalization of marketing from the aspects of business model innovation, precision marketing and data security, and studies the digital practice research of online network, product tracking and user information management, so as to improve the digital network of deep integration among enterprises, products and consumers.

Machine learning has a crucial role in people's lives. Machine learning can be divided into four parts: supervised learning, unsupervised learning, semi-supervised learning and reinforcement learning. It can help people with image recognition, stock prediction, etc. In addition, machine learning has a subset called deep learning (DL). DL in machine learning can help people optimize and learn models better. DL in machine learning can help people optimize and learn models better. One of them is generative adversarial networks (GANs). The other one is Convolutional Neural Networks (CNNs). First, GANs have achieved impressive results in image generation, data enhancement, etc. However, effective intervention and control of the results of generative adversarial networks is a challenging problem. Based on grayscale image data, this paper investigates the change patterns of the generation results of generative adversarial networks in the face of different input data. There are three groups in the experiment. The first group is to explore the effect of the size of the dataset on the generated images. The second group is to explore the effect of the diversity of the book dataset on the results. The third group was to explore the effect of the overall dataset on the generative images. The experimentally generated images revealed that data size and diversity are important factors that affect the quality of the results generated by generative adversarial networks. Among other things, the increase in the amount of data does not always positively affect the generative results.

Although the schemes based on deep learning applied to car autonomous driving show good performance in object detection un- der condition of normal weather, there are many difficulties in complex weather scenarios. When facing heavy fog interference, the general solution is to design two networks for image enhancement and object detection respectively, which means one dehazes the foggy image and the other detects the dehazed image. But this may bring two problems: 1. After the image is dehazed, it may cause some distortion of the image, especially the areas that are unlikely to be reconstructed due to severe haze; 2. For object detection, the interesting part of an image is the nearby area where the object is located. The rest background area that is far away from the object may not help. In response to the above two problems, in this work we proposed to use foggy images except whose area of object bounding box is clear as ground-truth in the training of image enhancement. In other words, only the key area around the object is clear in order to guide the network to perform image dehazing operations focusing on the target areas and finally improves the following object detection performance. By using U-net for image dehazing on the Foggy Cityscapes dataset and Faster R- CNN network with the same structure for object detection, the average accuracy mAP is increased by % compared to normal training, which demonstrates the effectiveness of our approach on foggy image object detection.

This paper proposes improved low-light net (LLNet) parameters based on the imaging features of optical remote sensing satellites, which are built on the conventional LLNet low-light image enhancement network. The network can be better adapted to the characteristics of thin cloud optical remote sensing satellite images by using the improved network parameters for denoising auto-encoder (DA) layer design, data training, and improvement effect evaluation of LLNet. This will enable the network to extract data features from the images and carry out adaptive image enhancement without over amplifying the highlighted areas. A set of specially produced high-definition satellite images is used as the experimental data set to validate the network parameter design experimentally, and the results show that the proposed improved network parameters can effectively enhance the simulated and real thin cloud optical remote sensing satellite images and can eliminate/diminish the serious influence of thin clouds on the features to a certain extent. After the enhancement of the simulated thin cloud image, its structural similarity with the reference image is improved by about 300%. Its subjective enhancement effect is more obvious even for the real thin cloud image.

5G gateway is an important equipment to solve the requirements of high-speed data transmission and low delay in industrial applications. For example, in factory workshops, 5G gateways are not only used for mobile devices such as AGV, but also for large-scale data transmission and real-time data analysis. In large enclosed industrial places, such as ports and mines, 5G gateways are used for remote control and other operations. In open scenarios such as natural gas pipelines, transmission lines and rivers, 5G gateways are used to realize remote inspection and monitoring. The industrial 5G gateway provides many key capabilities to support industrial applications, including providing equipment access capabilities, data collection capabilities, edge computing capabilities, data processing capabilities, protocol conversion capabilities, data forwarding capabilities, remote configuration management capabilities, security encryption protection capabilities, and anti-interference capabilities. This paper introduces a high-performance 5G industrial smart gateway hardware design scheme to meet the requirements of industrial applications. The actual test shows that the performance of this design scheme on 5G high-speed transmission meets the needs of industrial applications.

Generative Adversarial Networks have been widely used in recent years. Some extensions, such as Deep Convolutional Generative Adversarial Networks have also appeared. Currently, there is no research on the results generated by different pairings of existing generators and discriminators. Therefore, the research topic of this paper is to find the pairing combination of generator and discriminator to generate the best images. The research methods of this paper are as follows: Based on the Celeb-A Faces face dataset, the generated results were compared and analyzed by changing the generator and discriminator and considering the influence of the number of training epochs on the results to find the optimal combination. This paper has chosen two convolutional neural network generators, and discriminators, each of them has 3 and 5 layers of convolutional layers and one multilayer perceptron generator and discriminator. It is found that the generator and discriminator with higher convolutional layers generate clearer images. Multilayer perceptron generators and discriminators produce poor quality results. The larger the number of training epochs is, the better the generation effect is within a specific range. Therefore, selecting generators and discriminators with high convolutional layers is suggested for training.

Data augmentation by using Semi-Supervised Learning is an important research direction currently. Usually, this method applies the possible score of prediction as the confidence degree for the next training iteration so that Semi-Supervised Learning may gradually enhance the model's accuracy. The principle of Semi-Supervised Learning is using the original model to predict the unlabeled data to produce pseudo labels, then choosing the high confidence pseudo label to train the model iteratively. However, this method may not be helpful due to the iterative steps. This paper uses the model in this iteration to predict the data labeled in the last iteration and compares the pseudo label identity with the result produced by the previous iteration model. The evaluation shows that the consistency between models in two consecutive iterations is not high, which explains that Semi-Supervised Learning does not eventually enhance the model's accuracy. This research indicates that applying consistency and accuracy jointly as the standard for Semi-Supervised Learning in sentence sentiment analysis is a more reasonable method.

The ELT (Emergency Locator Transmitter) is a device used to locate an aircraft in an emergency situation, such as an accident, in order to facilitate search and rescue efforts. Conventional ELT devices have two triggering methods, namely automatic triggering when crash acceleration is reached and manual triggering when a button on the control panel is pressed. This paper proposes an ELT triggering device based on fingerprint collection technology. By integrating the fingerprint collection device into the control button and cross-linking it with the on-board data link, the fingerprint of the triggering person is collected, stored, and compared, so as to make a judgement or seek evidence, which effectively solves the problem of false triggering of traditional triggering devices and provides a basis for subsequent investigation of emergency situations. Experiments and use case study validate the usefulness of this new device.

Lyric transcription is similar to speech recognition, both identify content from sound clips. Speech recognition technology is maturing and related application systems have been widely used in the software industry, but the research on singing content is far from getting enough attention, there is still little research on identifying words and sentences from singing voice. What's more serious is that compared with the lyrics transcription in the English field, there are almost no related academic papers in the Mandarin field. On the one hand, speech recognition has high-quality datasets in multiple languages that are large enough to train large-scale models. However, the field of singing lacks data resources. On the other hand, compared with speech recognition, singing recognition has obvious skills in pronunciation, which is embodied in musical characteristics such as pitch and rhythm. Based on these problems, this paper aims to provide a dataset that can be used for Mandarin lyrics transcription, and build a transcription model on this dataset. Our model can address some deficiencies of the existing models, and achieves promising results on our dataset.

Bouyei Opera is one of the Nuo operas performed by the Bouyei people, the main ethnic group in Libo County, Guizhou Province. With a long history, Bouyei Opera has become an intangible cultural heritage (ICH) in China. However, the current development of Bouyei Opera is facing challenges in terms of talent cultivation, conservation of cultural heritage, as well as cultural promotion. Nevertheless, the advancement of artificial intelligence and information technology provide necessary technical support for the collection and promotion of China’s ICHs. In this paper, we first focused on collecting and classifying various Nuo operas in China. To address the enormous number of intangible cultural heritages and the disadvantages of relying on human resources to classify them, we utilized deep learning (DL) to annotate images and videos automatically. Specifically, we adopted clustering methods and a deep convolutional neural network (DCNN)- based unsupervised learning model to extract the semantic features for classifying all the ICHs. Then we obtained six types of databases covering the history, performance, style of singing, masks, cultural continuity, and troupes of Nuo operas, and designed web pages based on JavaScript frameworks according to these databases. We aim to provide a better visual experience to attract the public’s interest in Nuo operas. Finally, we used augmented reality (AR) technology to enhance the interaction at the terminal and offer a platform for the public to gain a broader understanding of non-heritage cultures and skills, thereby increasing the visibility of non-heritage products.

Machine learning is an important research area belonging to artificial intelligence technology. Machine learning has potent data processing and prediction ability. This paper uses machine learning to build a data analysis model and validate it with the NBA. This paper consists of two parts in total. The first part uses a clustering algorithm to analyze all teams' high-order data from 2000 to 2021 and classify the teams' styles. The main effect of using the k-means algorithm is to put the data in very different clusters, and in this paper the different clusters are considered to be the different team styles. At the same time, we can determine whether the style of the team has changed. We used the Euclidean distance as the distance formula for k-means. By the test of different k numbers, the result showed that between the numbers 2 to 9, the Silhouette Coefficient performed best when and second when. This paper used to measure the method that can be used to evaluate the influence of different algorithms or operating modes of algorithms on clustering results based on the same original data. For example, when k=2 and k=4. The experimental results show that the team styles in the NBA can be broadly classified into several categories over a long development period. The second part constructs a win prediction model based on the Elo method for the matchup data from 2020 to 2022. To sum up, in this paper, Elo ranking system, the Logistic regression model and three data sets are used to get a detailed prediction of the winning rate of NBA teams in 21-22 seasons.

Pedestrian detection in infrared images has been a hot and difficult research topic in computer version. Traditional methods of pedestrian detection mainly depend on the manual feature for the expression of human body and the results largely relies on the feature representation. Designing artificial features is time-consuming and labor intensive, requires heuristic expertise and experience. Deep learning model based on convolution neural network can automatically learn feature representation from the original images, while avoiding the drawbacks of artificial features. Its difficulty is the choice of network parameters. In this paper, we propose to use deep learning method based on convolution neural network in the process of pedestrian detection. In addition, we analyze the impact of network layers, convolution kernel sizes and feature maps to pedestrian detection in infrared images. The results demonstrate the superiority of our method over traditional methods in detection rate and alarm rate.

The data parallelism provides greater efficiency under multiple-node systems. Under these circumstances, there is more and more utilisation on multiple GPUs for better efficiency for computation and lower cost of time for accomplishing the project. However, some typical, ordinary, and common circumstances in which using multiple GPUs is worse than using one GPU in a machine. This paper aims to prove that using various GPUs is not omnipotent in efficiency and cost of time when running a program. This paper represents the limitations of computation with CPU, GPU, and multiple GPUs. Two main parts represent the limitations of the experiment. They are the comparison between one running CPU and one running GPU and the comparison between one running GPU and running multiple GPUs with the cost of time after running a program by numerous times and different data sizes

Facial expression is a part of body language, which can often convey physiological and psychological reactions. Facial expression has always been a hot topic in the area of image analysis and pattern recognition, whose basic task is to model facial images to recognize human emotions at a certain time. Due to CNN’s strong ability on feature expression, the researches of facial expression recognition based on deep learning have developed rapidly. In this paper, in order to improve the accuracy of facial expression recognition which is low because of the complex environments and structure of the existing expression recognition network, we propose a facial expression recognition method based on region of interest extraction, which uses ROI pooling to extract the local features of different face regions to construct a fixed size shared feature layer. In addition, we also add L2 regularization and learning rate decay mechanisms to find the optimal solution. Quantitative and qualitative experimental outcomes show that it is an effective way to achieve the facial expression task.

Object detection is a popular direction in the field of computer research, and the main task is to accurately classify and localize object instances in images. Since the advent of deep learning technology, it has advanced significantly in both speed and accuracy. However, there is still potential for improvement in the areas of multi-scale object detection, weakly supervised object detection, and unsupervised object detection. In this research, we examine and describe the current mainstream object identification algorithms based on deep learning from the viewpoints of process, network structure, common data sets, and model training approaches, based on detailed reference to a vast quantity of existing literature. We also give improvement suggestions according to the main problems and look forward to the future research focus in the field of object detection.

After years of research on traffic accidents, it has been proved that serious traffic accidents will be caused by drivers' poor state, inattention, fatigue and other conditions. Computer vision obtains the driver's body information and face information through visual recognition technology, compares and analyzes with a large number of face database information, judges the driver's state, and makes corresponding reminders and warnings, so as to achieve driving safety. In this paper, the infrared camera is used for real-time face recognition and key point detection, and a multi feature fatigue driving detection method is proposed. Combined with OpenCV, the eye, mouth and head spatial posture coordinate points of the human face are located, and the fatigue is determined according to the change degree of blinking, yawning and nodding. Finally, the fusion algorithm is used to synthesize the above fatigue characteristic factors for fatigue prediction. Experiments show that the accuracy of the algorithm is more than 96%, and it has good stability and anti-interference ability.

Current ECG classification models focus on the performance of the classification and do not focus on the interpretability of the classification results. This paper proposes an interpretation method for ECG classification results based on deep learning. This method determines the key heartbeat and key ECG time by replacing the heartbeat with a normal heartbeat, setting the fixed-width ECG data segment to zero, and analyzing the changes in the classification result. The classification contribution value of the segment to the classification result, and the heartbeat and electrocardiogram time segment with a larger contribution value to the classification result become the key heartbeat and key time segment for the classification. The experimental results show that the etiological explanation established by this method is highly consistent with the doctor's explanation, which partially solves the interpretation problem of the ECG classification results based on deep learning.

Bitcoin is a currency with high price volatility, and it is difficult to predict the Bitcoin price accurately. In recent years, many optimization strategies based on neural networks have been proposed to improve the performance of prediction tasks. The most common ones are extending the dataset, adding extra features, and adjusting output length. This paper aims to verify whether these strategies work on the Bitcoin prediction problem. Three relevant experiments are proposed to train gated recurrent unit models, one of the recurrent neural network models. The evaluation scores are calculated using mean absolute percentage error, root mean square error, and the coefficient of determination. The results show that none of the above three methods can improve the model performance on a four-year daily trading dataset, which proves that the Bitcoin market has its characteristics. In addition, the results demonstrate that R-squared is necessary for evaluating the result of the regression problems. Finally, this paper suggests designing an evaluation function for the Bitcoin price prediction problem to meet the prediction purpose better: reduce investment risks.

Various emergencies occur frequently, posing threats and challenges to people’s lives and social security. In consequence, the evacuation of multi-Agent has become a significant part of the emergency response process. However, a few existing works only focus on the evacuation of a small number of agents, which does not consider the problem of multi-Agent cooperation caused by the increase of the number of agents and the impact of emergencies. Therefore, a framework for event-driven multi-Agent evacuation is proposed in this paper, which includes three parts: event collection, event sending, and task execution. During task execution, agents are divided into groups and select the leader in the group, while other agents in the group move with the leader. Then, the reinforcement learning algorithm Space Multi-Agent Deep Deterministic Policy Gradient (SMADDPG), proposed in this paper, is used for path planning. In addition, the state, action and reward based on the Markov game are designed, and an environment with emergencies is presented as agents evacuation scenario. The experiment results show that the method proposed can shorten the length of path, and improve the interoperability between multi-Agent when emergencies occur, which can provide decision-making reference for emergency departments to formulate evacuation plans.

This article provides a brief overview of a few recent traffic sign detection studies, as well as a review of the concept and structure of traffic sign detection throughout the last decade. The methodology differs in several ways, but it is commonly divided into four distinct aspects. They are methods that are based on color or shape, hybrid-based methods, and methods that are based on machine learning, in that order. Machine learning-based technologies have steadily seized the lead recently because of their superior performance. As a result, the main attention of this study is on machine learning methods, as well as a review of past work, datasets, and performance. The outcomes of experiments using the same methodologies but different approaches are compared.

This paper is a sample of machine learning algorithms applied to analyze how different factors affect the game's result among all National Basketball Association teams. To achieve the study's objectives, this paper selects cluster analysis methods in supervised and unsupervised learning, such as the support vector machines, the nearest-neighbor classification, and the k-means algorithm to build the model. The first step is to classify the training and test sets, to calculate the accuracy and thus obtain the most efficient method of distinguishing between strong and weak teams, and subsequently to visualize the results by downscaling the influencing factors to find the main influence factors that differentiate between strong and weak teams. Finally, the cluster analysis led to the conclusion that the degree to which a team's playing style fits the league's seasonal rules and the values of the variables of the basic data is not directly related to the team's win rate.

Numerous network models can successfully fit some datasets since the advent of deep learning. The performance of the deep learning model continues to improve, mainly benefiting from the constantly added modules, such as convolution kernels. To quantitatively evaluate the role of these modules, it is essential to research how each module in the network topology affects the network's functionality by developing the standard model in this study. The paper takes LeNet, AlexNet, and VGG19 as examples, considering how each module affects prediction accuracy. In LeNet, AlexNet, or VGG19, different outcomes can be obtained by altering the linear layer of the net's dimension, the convolution kernel size, the dropout layer's percentage, or the pooling layer's deletion. This experiment is based on the CIFAR-10 dataset to test the original network. According to the experimental findings, the network's performance will also depend on the size of the convolution kernel, the dropout ratio, and whether the pooling layer is kept.

Side-channel attack is a commonly used attack method for recovering cryptographic chip keys, and plays an important role in the field of cryptographic chip physical security evaluation. Combining side-channel attacks with machine learning and replacing some steps of traditional side-channel attacks with machine learning methods can improve the efficiency of key-recovery from side-channel attacks to a certain extent. In practice, there is a problem that most existing cryptographic chip security evaluation systems cannot support the complete key recovery process, and fully improve the utilization of side information generated in the evaluation process. In this paper, we design a cryptographic chip physical security evaluation system based on machine learning. Through the integrated operation of power trace acquisition, preprocessing, analysis and evaluation, the correct key can be successfully recovered.

As a pillar industry in China, real estate has provided great help for economic development. At the same time, buying a house has become a topic everyone cannot avoid. Until now, many scientific literatures and materials have proved that gradient boosting is a basic strategy. In this paper, the authors first clean the captured data and select features. Then, three methods eXtreme Gradient Boosting (XGBoost), Random Forest, and Bi-directional Long Short-Term Memory (BiLSTM), are used to predict housing prices in Beijing, one of the most representative cities in China, to investigate and compare the efficiency gradient methods of the three methods. The above three models process and predict 23 factors that affect housing prices by collecting information on houses sold and sold in Beijing in recent years. The score curve is obtained by fitting, finally, the best model of prediction is selected by adjusting the parameters. The influence of different characteristics on housing price prediction is studied. The experimental results show that compared with XGBoost and random forest, BiLSTM has a greater advantage in forecast speed and accuracy and is the most suitable model for predicting housing prices.

With the rapid development of authenticated encryption (AE) algorithms in recent years, especially after the CAESAR (Competition for Authenticated Encryption: Security, Applicability and Robustness) competition was launched, a large number of excellent authenticated encryption algorithms have emerged, making the analysis of authenticated encryption algorithms a hot research issue. The CAESAR competition was launched under the sponsorship of IACR in 2014, aiming to collect excellent authentication encryption algorithms from all over the world. SAEAES is a relatively excellent authentication encryption algorithm in the CAESAR competition. In this paper, the sponge structure of SAEAES is improved by introducing the MD (Merkle Damgard) iterative structure. At the same time, in order to improve the ability to resist collision attacks, fixed point attacks and cluster attacks, the MD iterative structure itself is also improved.

The significant success of Deep Neural Networks (DNNs) relies on the availability of annotated large-scale datasets. However, it is time-consuming and expensive to obtain the available annotated datasets of huge size, which hinders the development of DNNs. In this paper, a novel two-stage framework is proposed for learning with noisy labels, called Two-Stage Sample selection and Semi-supervised learning Network (TSS-Net). It combines sample selection with semi-supervised learning. The first stage divides the noisy samples from the clean samples using cyclic training. The second stage uses the noisy samples as unlabeled data and the clean samples as labelled data for semi-supervised learning. Unlike previous approaches, TSS-Net does not require specifically designed robust loss functions and complex networks. It achieves decoupling of the two stages, which means that each stage can be replaced with a superior method to achieve better results, and this improves the inclusiveness of the network. Our experiments are conducted on several benchmark datasets in different settings. The experimental results demonstrate that TSS-Net outperforms many state-of-the-art methods.

Clustering is an important part of artificial intelligence and is widely used in data mining, pattern recognition, natural language processing, computer vision and other aspects. Clustering algorithms are complex and difficult to master. In order to make learning easier and more interesting, a visual clustering experiment is designed. The first is the introduction of clustering algorithms and clustering indicators; then basic experiments are designed to show the working principle and clustering characteristics of each algorithm through the visualization of the clustering process and the visualization of the clustering results, and introduce the self-generated artificial data to the method to verify the characteristics of the algorithm; the last is the application experiment, through the face clustering and image segmentation experiments to improve the students' interest in learning.

With the rapid development of deep learning, UAV target detection technology based on computer vision and artificial intelligence has been widely used in practice. However, due to the instability of UAV movement, limited by load and endurance, the development of UAV target detection is slow, and there are challenges such as significant changes in target scale, occlusion between objects, and changes in target density. This paper builds on the network model structure of YOLOv5 to address these challenges. It adds a detection head generated from low-level feature layers and high-resolution combined feature maps to detect tiny objects. We utilize the Bifpn network structure and a weighted fusion splicing approach to fuse more features and introduce an improved Coordinate Attention to obtain location information for feature enhancement accurately. Extensive experiments on the Visdrone2021 dataset show that the model achieves good results in UAV target detection and is helpful for tiny and occluded target detection.

Nowadays the space borne synthetic aperture radar (SAR) is an important tool to acquire the information of the ground features. As is well known, SAR image is a typical High Dynamic Range Image (HDR). HDR is generally beyond the dynamic display range of the commonly-used display equipment, and visual interpretation is the most common and direct means for researchers to obtain ground object information from SAR images. Therefore, in order to obtain high quality SAR image, it is necessary to map high dynamic range (HDR) image into low dynamic range (LDR) display output. In this article, we will analyze of SAR image brightness characteristics and propose a segmented mapping algorithm suitable for SAR image. In addition common image quality evaluation index are described.

The COVID-19 epidemic has spread throughout the world and poses a serious threat to human health. Any technical device that provides the accurate and rapid automated diagnosis of COVID-19 can be extremely beneficial to healthcare providers. A new workflow for performing automated diagnosis is proposed in this paper. The proposed methods are built on a well-designed framework, two kinds of CNN architectures including a custom CNN and a pre-trained CNN are utilized to verify the effectiveness of the focal loss function. According to the experimental findings, both CNNs that were enhanced with the focal loss function converged faster and achieved higher accuracy on the test set, outperformed the models that utilized cross-entropy loss that does not consider the class-imbalanced issue in the multi-class image classification with imbalanced Chest X-ray (CXR) image datasets. In addition, image enhancement techniques turned out to be very helpful for enhancing the CXR image signatures to achieve better performance in our work.

At present, the development of artificial intelligence is very rapid, and the intelligent assisted driving system based on deep learning is widely used in the society. For example, in unmanned driving, it can accurately identify pedestrians, vehicles and traffic signs. Convolutional neural network in deep learning has excellent achievements in the field of computer vision and has outstanding feature extraction ability. Therefore, object detection algorithm based on deep learning is a research hotspot in the field of computer vision at present. We propose a vehicle-pedestrian target detection method based on Yolov4-tiny. Firstly, the ResBlock-D module in the ResNet-D network is used to replace one CSPBlock module in Yolov4- tiny, thus reducing the computational complexity. Then, the coordinate attention mechanism is added to help the model better locate and identify targets. Experimental results show that The improved Yolov4-tiny algorithm has higher curacy than the original algorithm, and the Map is improved by 7.8 %, which has a certain reference value for the study of intelligent assisted driving technology.

As the most-traded digital asset, Bitcoin receives a tremendous increase in investment interests. The primary purpose of this paper is to examine and compare two commonly applied machine learning algorithms on their capability and feasibility in Bitcoin price prediction. The regression models of Extreme Gradient Boosting and Long Short-Term Memory are selected as the investigated objects in this paper. The experiments are evaluated by considering the extra impacts of sample dimensions and time-series frequency, along with the trade-offs between prediction accuracy (measured by residual error) and computational efficiency (measured by computing time). Long Short-Term Memory, as a more convoluted deep learning method, achieves better accuracy when a daily dataset with limited input features is used. However, its predictability has considerably decreased and thus become less efficient than XGBoost, when more peripheral information and higher frequency data points (trading price every 15 minutes) are available. Besides algorithmic complexity, Long Short-Term Memory also takes much longer computing time than Extreme Gradient Boosting, making it a less applicable model to use when dealing with large sample sizes.

Blockchain Embedded Architecture Protocol network is affected by homomorphic disturbance of nodes in routing and forwarding control, and is vulnerable to embedded intrusion. In order to improve the security of the network, an intrusion detection algorithm based on deep residual learning for Blockchain Embedded Architecture Protocol network is proposed. The statistical feature model of blockchain embedded architecture protocol network intrusion is constructed. The edge information fusion technology of association rules is used to detect and filter the intrusion features of blockchain embedded architecture protocol network, and the spectral features of blockchain embedded architecture protocol network data are extracted. The depth residual learning method is used to carry out the adaptive optimization control in the intrusion detection process of blockchain embedded architecture protocol network, and the multi-threshold depth residual fusion method is used to realize the accurate detection of abnormal traffic data and improve the ability to accurately locate and detect intrusions. The simulation results show that this method has a high accuracy probability, good detection performance and improved network security.

The changes in temperature may arise risks in many industries. To solve this problem, the National Meteorological Center and Dalian Commodity Exchange jointly compiled a temperature index which includes 5 cities. Therefore, forecasting time series temperature data in those cities is an important subject. Traditionally, we use statistic method ARIMA to predict the next lags of time series. With the advancement in computational power of computers and the introduction of more advanced machine learning algorithms, this paper develops a method by integrating ARIMA with machine learning to analyze and forecast time series data. The empirical studies conducted show that integrating ARIMA with Long Short-Term Memory outperforms that with Support Vector Regression, or Random Forest in their prediction accuracy.

Fish counting is one of the key issues in fish farming and trade. However, manual counting is not only expensive and inefficient, but it can also harm the fish. In order to solve the above problems, this paper proposes a counting model named VSPNet based on convolutional neural network to realize intelligent counting of snakehead fish. Firstly, the snakehead counting dataset is created, and the snakehead objects in each image in the dataset are labeled to obtain the true density map; and then the EESP module is added after the 10th convolutional layer of the VGG model to increase the receptive field, extract deep-level semantic information, and generate the estimated density map; finally, the total number of snakehead in the image is predicted from the density map. The experimental results show that the method proposed in this paper outperforms the MPS and DSNet models on the snakehead fish counting dataset, which proves the effectiveness of the method.

Medical image analysis based on computer vision technology has always been a research hotspot in the community, which aims to assist doctors in diagnosis by accurately analyzing pathological images to divide the patient’s condition and the patient’s lesions. Thanks to the rapid development of deep learning, the application of computer image recognition technology in medicine is becoming more and more widespread, while still facing a series of challenges such as low data set data, insufficient performance of algorithms and fine delineation of lesions. In order to solve these problems, based on extensive literature research, this paper first compares the algorithms in the application for Corona Virus Disease 2019, skin cancer and liver cancer. Then we introduce the improvement of these algorithms by expanding the number of data sets, optimizing the algorithms, and fitting the neural networks and models, whcih can improve the accuracy of image recognition technology to assist doctors in identifying lesions in clinical practice. The algorithms are further compared quantitatively on the basis of the training set in clinical diseases, and the difficulties to be overcome in image recognition and the future development trend are explained and predicted from the analysis of the comparison. Many new algorithms and excellent models are being gradually improved with the development of the times, and image recognition technology will also develop towards more research fields in the future.

Falls is one of sudden, unintentional changes in body position, which usually occur in the elderly and may seriously affect the physical and mental health of the elderly. Thanks to the rapid development of computer vision algorithms, fall detection research based on RGB images or videos has gradually become the mainstream framework for fall detection due to its rich semantics, low cost, and friendly user experience. In this paper, as a means of improving the accuracy of fall detection in real time, a lightweight fall detection method based on Lightweight OpenPose is proposed. Specifically, the proposed method first calculates the skeleton map and joint point coordinates of the human body in real time based on the Lightweight OpenPose. Then, the obtained coordinates are combined with the proposed fall detection algorithm for detection. Through extensive experiments, we qualitatively and quantitatively verify the effectiveness of our method. At the end of this paper, we also put forward some views on the further improvement measures of this algorithm.

With the Images from large telescopes accumulating, more and more unknown celestial bodies are being discovered by humans. However, determining whether an object is a star or a galaxy through manual methods is often time-consuming and inaccurate. Therefore, this paper collects celestial images obtained by astronomical telescopes from SDSS and classifies them using a Convolutional Neural Network. It is clear that when the image is clear, the accuracy of the model on the test set can reach more than 98%, and it can complete the classification well. Furthermore, the performance of the model under noise disturbance is tested for many times and its robustness is found to be poor. Under the attack of Fast Gradient Sign Method, the classification accuracy of the model is relatively lower than expected, and the anti-interference ability is poor, so optimization measures need to be implemented. After adding noisy images to the dataset, the model was reconstructed and retrained, which improved the classification accuracy of the model. The results show that when there are many images with noise in the test set, the accuracy can reach about 86%, which is proved to be an effective means to defend FGSM attacks.

A motor imagery brain-computer interface system with practical application value should be able to show stable performance when facing new users. The distribution of electrodes on the cerebral cortex is the same for any user. Therefore, in order to solve the subject-independent problem, we propose a novel Graph Convolutional Convolution Transformer Net (GCCTN), which uses a graph convolutional neural network to calculate the relationship between an electrode and other electrodes, uses a convolutional neural network to extract temporal and spatial information and uses a Transformer Encoder for further extraction of time-domain information. Finally, the classification accuracy of our model is optimal.

The big data of mobile resources under the new computing network architecture is repetitive and redundant, which leads to poor classification in the process of data scheduling and detection. In order to reduce the error rate of big data deduplication and unloading of mobile resources under the new computing network architecture, a new method of big data deduplication and unloading of mobile resources under the new computing network architecture based on redundant data elimination is proposed. Autocorrelation matched filter detection model is used to filter redundant data and suppress symbol interval interference on the prior features of mobile resource big data under the new computing network architecture with random sampling, and the clustering convergence characteristic parameters of mobile resource big data under the new computing network architecture are extracted by using sample fuzzy regression analysis and least squares sample block fusion detection method. The constrained evolution method of multi-level iterative regression analysis is used to estimate the classification features of mobile resources big data under the new computing network framework. The classification target features are input into the BP neural network classifier, and the adaptive weight distribution control of BP neural network classification is carried out by combining the adaptive clustering center optimization control algorithm, which improves the adaptability of data classification and realizes the unloading of mobile resources big data under the new computing network framework. The simulation results show that the algorithm can effectively reduce the interference of redundant data, and the fidelity rate of data classification is high and the error rate is low, which improves the dynamic management ability of mobile resource data under the new computing network architecture.

With the growth of the used car market and the development of e-commerce platforms, the need for accurate valuation of used car prices is becoming more urgent. Accuracy of price evaluation is the key to the success of used car transactions. At present, the common methods are manual experience method, Monte Carlo method, etc. Among them, manual experience method and multi-attribute decision method are more mature and widely used in traditional pricing method, but they have some disadvantages such as large computation amount and low accuracy. Aiming at the above problems, a BP neural network model based on mean encoding is designed in this paper. After extracting the features of the model, BP neural network is used to study the pre-processing data to predict the price for the network output. In this paper, a real used car trading dataset was used to test the model. The 2 R error is 0.976. Compared with the SVM and the decision tree model, this model is more accurate.