This PDF file contains the front matter associated with SPIE Proceedings Volume 6788, including the Title Page, Copyright information, Table of Contents, Preface, and the Conference Committee listing.

Automatic affective expression recognition has attracted more and more attention of researchers from different disciplines, which will significantly contribute to a new paradigm for human computer interaction (affect-sensitive interfaces, socially intelligent environments) and advance the research in the affect-related fields including psychology, psychiatry, and education. Multimodal information integration is a process that enables human to assess affective states robustly and flexibly. In order to understand the richness and subtleness of human emotion behavior, the computer should be able to integrate information from multiple sensors. We introduce in this paper our efforts toward machine understanding of audio-visual affective behavior, based on both deliberate and spontaneous displays. Some promising methods are presented to integrate information from both audio and visual modalities. Our experiments show the advantage of audio-visual fusion in affective expression recognition over audio-only or visual-only approaches.

Combining different modalities for pattern recognition task is a very promising field. Basically, human always fuse information from different modalities to recognize object and perform inference, etc. Audio-Visual gender recognition is one of the most common task in human social communication. Human can identify the gender by facial appearance, by speech and also by body gait. Indeed, human gender recognition is a multi-modal data acquisition and processing procedure. However, computational multimodal gender recognition has not been extensively investigated in the literature. In this paper, speech and facial image are fused to perform a mutli-modal gender recognition for exploring the improvement of combining different modalities.

In this paper, an efficient rotation invariant texture classification method is proposed. Comparing with the previous texture classification method, which is also based on Gabor wavelets, two modifications are made in this paper. Firstly, an adaptive circular orientation normalization scheme is proposed. Because both the effects of orientation and frequency to Gabor features are considered, our method can effectively eliminate the disturbance from inter-frequency, and therefore has the ability to reduce the effect of image rotation. Secondly, besides the Gabor features, which mainly represent the local texture information of an image, the statistical property of the intensity values of an image is also used for texture classification in our algorithm. Our method is evaluated based on the Brodatz album, and the experimental results show that it outperforms the traditional algorithms.

Radar target's HRRP always has some information redundancy, and is easily to be affected by noise or lack of separability. In this paper, using the advantage of kernel methods for solving nonlinear forms, we propose a radar target's HRRP feature extraction method based on Kernel Principal Component Analysis (KPCA) and a radar target fuzzy recognition method based on Support Vector Data Description (SVDD). In the course of feature extraction, KPCA method is used to reduce radar target's HRRP and to compress the dimension of HRRP, so that we can depress the noise and the sensitivity of target posture; in the course of recognition, we first find the smallest hyper-sphere including every class of training samples in feature space, then construct the fuzzy membership function according to the distance between every testing sample and the hyper-sphere surface, so we can recognize every testing sample based on its fuzzy membership. Simulation results of multi-target recognition reveal that the new method proposed in this paper not only achieves high recognition accuracy, but also has excellent generalization performance, for instance, we can achieve high recognition accuracy in lower SNR. So the new feature extraction and recognition method proposed in this paper is particularly suitable for radar target recognition.

Consensus building in group support systems relies on the mutual-question and mutual-elicitation of experts, so a feedback mechanism is required to conduct experts to converge their thinking by visualizing the individual opinion and the consistent state of the group. This paper proposes a new feedback mechanism, which first clusters the experts' preferences into a set of subgroups, and then uses different line-types or line-colors to display the clustered opinions in parallel coordinate. By using this mechanism, the group consistency is analyzed and the group discussion is conducted efficiently. One of the characteristics of the proposed method is that it can protect the minority views automatically. An example is presented to illustrate the application of the method.

This paper presents a novel scene classification method using low-level feature and intermediate feature. The purpose of the proposed method is to improve the performance of scene classification and reduce the labeled data required using the complementary information between low-level and intermediate feature. The proposed method uses the co-training algorithm to classify scenes, in which the low-level feature and intermediate feature are two views of co-training algorithm. For low-level feature, Block Based Gabor Texture (BBGT) feature is extracted to describe the texture property of images incorporating the spatial layout information. For intermediate feature, Bag Of Word (BOW) feature is extracted to describe the distribution of local semantic concepts in images based on quantized local descriptors. Experiment results show that this proposed method has satisfactory classification performances on a large set of 13 categories of complex scenes.

Put forward a method of 3D point reconstruction from an image taken by monocular camera. In algorithm, the coordinate conversion factor of the given point can be calculated through a pair of parallel lines, whose positions in the scene are known. Then reconstruction of any points in 3D space which contains parallel lines can be achieved using the coordinate conversion factor. In experiments, the algorithm is used to estimate the trajectory of walking people.

To reconstruct a 3D scene around the lunar rover from stereo image-pairs captured by the panoramic cameras, based on which an intuitive platform can be put up for scientists to plan exploration commands, we set about to study the 3D scene reconstruction. This paper mainly presents a scheme of registering local scene models to reconstruction a large scene. When a few of local 3D scene models have been reconstructed respectively, we firstly find the common 3D point sets between every two adjacent local models based on edge detection and image matching, secondly fit the matrix of coordinate transformation employing the technique of separating rotation matrix and translate vector, and lastly register these local 3D models into a uniform coordinate system. In this scheme, we don't need to set a few of control points in a scene beforehand; and we determine the rotation matrix by a system of linear equations based on Cayley transformation. Experimental results of reconstructing indoor and outdoor scenes show that our scene registration method is feasible.

In this paper we propose an approach to model the posterior probability output of multi-class SVMs. The sigmoid function is used to estimate the posterior probability output in binary classification. This approach modeling the posterior probability output of multi-class SVMs is achieved by directly solving the equations that are based on the combination of the probability outputs of binary classifiers using the Bayes's rule. The differences and different weights among these two-class SVM classifiers, based on the posterior probability, are considered and given for the combination of the probability outputs among these two-class SVM classifiers in this method. The comparative experiment results show that our method achieves the better classification precision and the better probability distribution of the posterior probability than the pairwise couping method and the Hastie's optimization method.

In this paper, we integrate a pair of CCD cameras and a digital pan/title of two degrees of freedom into a binocular stereo vision system, which simulates the panoramic cameras system of the lunar rover. The constraints for placement and parameters choice of the stereo cameras pair are proposed based on science objective of Chang'e-IImission. And then these constraints are applied to our binocular stereo vision system and analyzed the location precise of it. Simulation and experimental result confirm the constraints proposed and the analysis of the location precise.

A novel shape classification method based on Hidden Markov Models (HMMs) is proposed in the paper. Instead of characterizing points along an object contour, our method employs HMMs to model the relationship among structural segments of the contour. Firstly, an object contour is partitioned into segments at points with zero curvature value. Secondly, each segment is represented with structural features. Finally, a HMMs is utilized to characterize the object contour by treating each segment as an observation of a hidden state. Promising experimental results obtained on two popular shape datasets demonstrate that the proposed method is efficient in classifying shapes, particularly unclosed shapes and similar shapes.

Nowadays there are more and more targets, so it is more difficult for radar networking to track the important targets. To reduce the pressure on radar networking and the waste of ammunition, it is very necessary for radar networking to recognize the targets. Two target recognition approaches of radar networking based on fuzzy mathematics are proposed in this paper, which are multi-level fuzzy synthetical evaluation technique and lattice approaching degree technique. By analyzing the principles, the application techniques are given, the merits and shortcomings are also analyzed, and applying environments are advised. Another emphasis is the compare between the multiple mono-level fuzzy synthetical evaluation and the multi-level fuzzy synthetical evaluation, an instance is carried out to illuminate the problem, then the results are analyzed in theory, the conclusions are gotten which can be instructions for application in engineering.

In the industry, the three dimension of the object is often measured. But the method is usually the contact measurement. And the speed is slowly. The measurement is needed both high precision and fast speed, so the non-contact measurement is required. The grating projecting is the non-contact measurement with prospects. But there are some difficulties in the method. Firstly, when the object has the steps shape or there are shadows in the grating stripes, the disconnected phase can't be correctly unwrapped. Secondly, it is very difficulty to realize the real time digital filter. Now the digital filter is man-machine conversation, so the speed is slowly. Thirdly, in order to measurement the different object, the adaptive grating is needed. In order to resolve the above problems, the grating program is created on the computer. The program has many functions, including the phase shift, the two-frequency grating and the grating frequency is easy to adjust. So the adaptive grating is realized. The two-frequency grating is programmed by the computer. And it is projected to the measured object. The measurement object is placed on the exact rotary platform. The deformed grating is collected in the Charge Coupled Device (CCD). After getting two images, the two images are mosaiced. Then the clear object image modulated by the grating is got. The problem of the steps shape or there are shadows in the grating stripe is worked out. Then the fourier transform is used to process the image. In the traditional fourier transform profilometry, the phase is worked out as follows: After fourier transform, the zero frequency spectra is shifted to the origin of frequency, then filter the needed signal. Then the needed signal is shifted to the center of frequency, and then the zero frequency is shifted to both sides. After inverse fourier transform, the imaginary part is getting, so the phase is getting. But it has a difficult in the above method, because of three times frequency shift, and the center frequency is difficult to confirm, the frequency shift can't be correct and the filter can't be designed correctly, and the error can be transferred, so the result of filter is not well, it has bad effect to the later measurement. The result of measurement is also not well. In order to conquer the difficult, after the fourier transform, filtering the needed signal without frequency shifting, then inverse fourier transform. So the phase relational with the frequency and coordinate is getting. The phase of the reference surface is getting by the same method. Then the difference phase is getting. The real difference phase of low frequency is easy to got, then the real difference phase of high-frequency is work out based on it. At last, according to the relation of difference phase and the height, the three dimensional profilometry of the object is reconstructed. An example of step shape object is done. The three dimensional profilometry is reconstructed successfully. It takes 3 second to reconstruct the three dimensional profilometry. The precision is 0.5mm. The result indicates that the method has conquered the above problems. The result indicates that the method is simple, with fast speed and high precision. Three dimension profilometry measurement of the objected that have the step shape or the shadow in the projecting can be successfully resolved.

The rapid development of the astronomical observation has led to many large sky surveys such as SDSS, 2DF, LAMOST etc. Because of the sheer size of these surveys, it becomes urgent to develop methods of reliable and automated spectral recognition. A new cross correlation technique for redshift determination of galaxy spectra is presented in this paper. We use principle components analysis to construct galaxy templates. According to the redshift candidates determined by spectra line features, cross-correlation between the observed spectrum and the templates is measured by the weighted sum of several similarity evidences. The candidate of the highest correlation is chosen as the estimated redshift. Both simulated spectra and observed spectra are used to test the proposed method, the correct rate can reach 97% above.

According to the main characteristic of the city's water affairs dispatchment, the structure of water affairs dispatchment based on rough sets theory was proposed. After each factors were considered synthetically, knowledge expression system was set up, and the water affairs dispatchment control regulation was reduced and acquired. To some extent, it's a new method of processing the uncertain information in the water affairs dispatchment. The example demonstrates that this method has reduced the dispatchment control, and its regulation acquired is of objectivity, so it can solve preferably the control problem of the city's water affairs dispatchment.

Classification is a basic topic in data mining and pattern recognition. Following advances in computer science, a lot of new methods have been proposed in recent years, such as artificial neural networks, decision trees, fuzzy set and Bayesian Networks, etc. As a probabilistic network, Bayesian Networks is a powerful tool for handling uncertainty in data mining and many other domains. Naïve Bayes Classifier (NBC) is a simple and effective classification method, which is built on the assumption of conditional independence between the class attributes. This topology structure can not describe the inherent relation among the features. In this paper, we apply Bayesian Networks Augmented Naïve Bayes (BAN) for the texture classification of aerial images, which relaxes the independent assumption in NBC. A new method for learning the networks topology structure based on training samples is adopted in this paper. Comparison experiments show higher accuracy of BAN classifier than NBC. The results also show the potential applicability of the proposed method.

In LPR system, character recognition subsystem is heavily affected by image quality. To resolve this problem and improve recognition rate, a new algorithm is proposed, in which pulse coupled neural network (PCNN) is applied into the recognition of license plate character. PCNN model is simplified to improve computation efficiency, and then is utilized to extract three features from dimension-normalized binary result of input character image. Based on these features, weighted voting is performed and final estimation of input character is made. The experiment results show that compared with common algorithms based on BP network, the new algorithm based on simplified PCNN model has higher total recognition rate and stronger robustness, and is more convenient and flexible.

In order to settle the conflict of the visualization efficiency and the great amount data in digital city in 3d GIS system, the author put forward a 3D data model with LOD ability especially for the complex 3d object such as buildings in 3d city GIS system. In this paper, the author described the components of the complex 3d object, then explained that we can settle the conflict mentioned above by designing a new model that has the capability of describing levels of details. The author expound the basic theory of the model with LOD ability related with the vision point, and then defined several key conceptions, at same time, the author analyzed the principles and the visualization process of the model in 3d space. In the end of the paper, the author verified data model through an experiment, the results of the experiment showed that the model put forward is effective and high efficiency.

Introduce a CAD-based visual inspection system, which is designed to measure geometry dimension of sheet metal parts automatically, such as distance, angle, parameters of circle, etc. The inspection system extracts the features of sequence images depending on CAD data, and then to reconstruct real 3D of part. It outputs the measurement results against the various requirements of customer. The main contents of the paper include searching and matching of group of lines with image space constrains; visual inspection and reconstruction of sheet metal parts based on the photogrammetry generalized point photogrammetry. The result of experiments shows that the inspection system is robust and achieves the precision level of repeated manual measurement of an experienced inspector. The arithmetic discussed in the paper has potential to deal with another object with sharpness edges except sheet metal part.

With the emerging security demands, biometric identification technology has attracted more and more attention in recent years, and iris recognition is one of the most reliable biometric technologies. Iris localization is a crucial part in the iris recognition, which is quite time-consuming and easily disturbed by various noises, especially the eyelashes. A novel iris localization method is proposed in this paper. In the location of inner iris boundary, the gray curves of a row and a column with the pupil edge are used to estimate the coarse center and radius of pupil, which can reject the eyelash noises. The experiments show this coarse location method has better accuracy and speed than the common gray projection. Edge points of pupil are extracted by a gradient operator and fitted as the iris inner boundary. In the location of outer iris boundary, the image binarization is use to mark most noises, and then the outer iris boundary is extracted by integro-differential operator from the coarseness to fine. Performance experiments have been done, and the results show that about 0.175 second at speed and 99.5% at precision are reached by developed algorithm. In comparison with other classical methods, this algorithm has faster speed and better robustness.

Probability-Based Covering Algorithm (PBCA) is a new algorithm based on probability distribution. It uses the probability of samples and decides the class of the sample on the border of coverage by voting. In the original covering algorithm, there are many tested samples that can't be classified by the spherical neighborhood gained. The network structure of PBCA is mixed structure composed of feed-forward network and feedback network. The method of adding some samples of different class and enlarging the coverage radius is used to decrease the number of refused samples and improve the rates of recognition. The algorithm is effected in improving the study precision.

Differential Evolution (DE) method is introduced in this paper to make up the insufficiency of basic probabilistic neural network. Consequently, a new texture image recognition method based on Modified Probabilistic Neural Network (MPNN) is proposed. At first, tree structure wavelet packet transformation is used to extract the energy characteristic, and statistical method is used to extract the statistical mean value, average energy, standard deviation, and mean residual characteristics for obtaining the feature vector; then the feature vector of texture image is trained by the MPNN, thus the texture image is identified. The experiment result indicates that, compared to the BP neural network, RBF neural network, and the basic probabilistic neural network, the modified probabilistic neural network has higher accuracy and faster convergence speed.

In this paper, we proposed a manifold-based algorithm called Orthogonal Neighborhood Preserving Embedding (ONPE) for dimensionality reduction and feature extraction. ONPE algorithm is based on the Neighborhood Preserving Embedding (NPE) algorithm. NPE is an unsupervised dimensionality reduction method which is the linear approximation of classical nonlinear method. However, the feature vectors obtained by NPE are nonorthogonal. ONPE inherits NPE's neighborhood preserving property and produces orthogonal feature vectors. As orthogonal eigenvectors preserve the metric structure of the image space, the ONPE algorithm has more neighborhood preserving power and discriminating power than NPE. Furthermore, ONPE can find the mapping which best preserves the manifold's estimated intrinsic geometry structure in a linear sense. Experimental results show that ONPE is an effective method for feature extraction.

The development of hardware and software is not sufficient to meet the real-time visualization requirements of large scale 3D City Models. How to adaptively coordinate the speed and quality of rendering according to the data volume and hardware/software environment is therefore a critical issue. This paper proposes an algorithm which predicts the rendering time according to the features of 3D City Models at first, and then to calculate the object importance value based on the mathematical model which considers the indicators of each object: location, distance, visibility and semantics, and finally to select the object set to be rendered by a fast recursive algorithm. There are five factors selected to test their influence on rendering time: triangle number, vertex number, texture number, screen pixel number, and the texture image size. According to multivariate statistical theory, experimental results prove that both geometry and texture data size are significant for rendering time of 3D City Models. A typical 3D building group models are employed for experimental analysis. The results show that the method introduced in this paper is accurate to predict the time of rendering 3D models with detailed texture. The adaptive rendering performance is also significantly improved.

We propose a totally novel method based on a revised ant colony clustering algorithm (ACCA) to explore the topic of textural defect detection. In this algorithm, our efforts are mainly made on the definition of local irregularity measurement and the implementation of the revised ACCA. The local irregular measurement defined evaluates the local textural inconsistency of each pixel against their mini-environment. In our revised ACCA, the behaviors of each ant are divided into two steps: release pheromone and act. The quantity of pheromone released is proportional to the irregularity measurement; the actions of the ants to act next are chosen independently of each other in a stochastic way according to some evaluated heuristic knowledge. The independency of ants implies the inherent parallel computation architecture of this algorithm. We apply the proposed method in some typical textural images with defects. From the series of pheromone distribution map (PDM), it can be clearly seen that the pheromone distribution approaches the textual defects gradually. By some post-processing, the final distribution of pheromone can demonstrate the shape and area of the defects well.

Image semantic understanding is one of the most important techniques for solving the problem of semantic gap. By introducing generalized computing into image semantic understanding, this paper presents a kind of third class image description model. Then, under the guidance of the model, the approaches of image semantic information extraction is proposed based on generalized set and generalized transformation. Finally, a kind of image semantic understanding system based on generalized is sketched out.

We propose an appearance based model for face recognition in news videos using an enormously large databank of still images. This is a step towards building an elaborate face-query system using multimodal audio-visual data. We use the fact that faces of the same person appear similar than of different people. We preprocess the videos, apply feature extraction, feature matching and a unique parallel line matching algorithm to develop a simple yet a powerful face recognition system. We tested our approach on real world data and the results show good performance both for high resolution still images and low resolution news videos without involving any training or tasks like face rectification, warping etc. It can be incorporated as part of a larger multimodal news video analysis system with problems of time alignment between text and faces. Our results show that this simple approach also works well where video modality is the only source of information.

The simulation of ocean environment in the infrared has been a hot yet difficult problem in the field of computer simulation. In this paper, the shortage of the simulation of infrared ocean images with Vega is analyzed, and then a new simulation method based on 3D modeling with OpenGL is introduced. The new method abandons the high precision mesh but uses mathematical model to manipulate vertex of the mesh and establish the model. Experiments demonstrated that the method proposed is much more efficient and guarantees the quality of the simulation images. Finally a similarity evaluation function based on features extracted from co-occurrence matrix such as angular second moment, entropy, related coefficient, contrast and uniformity is put forward to evaluate the similarity of the images.

Logo recognition has gained much development in the document retrieval and shape analysis domain. As human computer interaction becomes more and more popular, the logo recognition through a web-camera is a promising technology in view of application. But for practical application, the study of logo recognition in real scene is much more difficult than the work in clear scene. To cope with the need, we make some improvements on conventional method. First, moment information is used to calculate the test image's orientation angle, which is used to normalize the test image. Second, the main structure of the test image, which is represented by lines patterns, is acquired and modified Hausdorff distance is employed to match the image and each of the existing templates. The proposed method, which is invariant to scale and rotation, gives good result and can work at real-time. The main contribution of this paper is that some improvements are introduced into the exiting recognition framework which performs much better than the original one. Besides, we have built a highly successful logo recognition system using our improved method.

Rapid texture mapping of buildings is a key aspect for reconstruction of 3D city landscapes. An effective approach by the way of coarse-to-fine 3D building model generation by integration of LIDAR and multiple overlap images is proposed. Classification and segmentation can be processed by combined multi-spectral information which is provided by color aerial image and geometric information from multi-return laser scanned data. A connected graph of the segment label image has to be created to derive the neighborhood relation of the planar segments. A line segment matching, based on geometry and chromatic constraint, is applied for automatically getting the corresponding line features in multi target images. Hypotheses for polyhedral surfaces are selected using topological relations and verified using geometry.

Nonlinear CCA extends the linear CCA in that it operates in the kernel space and thus implies the nonlinear combinations in the original space. This paper presents a classification method based on the kernel canonical correlation analysis (KCCA). We introduce the probabilistic label vectors (PLV) for a give pattern which extend the conventional concept of class label, and investigate the correlation between feature variables and PLV variables. A PLV predictor is presented based on KCCA, and then classification is performed on the predicted PLV. We formulate a frame for classification by integrating class information through PLV. Experimental results on Iris data set classification and facial expression recognition show the efficiencies of the proposed method.

Human tracking has attracted much attention from the researchers in the fields of computer vision and pattern recognition. The problem is generally extremely challenging partly because human bodies are articulated and versatile, and partly because background clutter, both of which demand a strong human model. However, there is usually a trade-off between the discriminative power and the complexity of a given model. This paper presents a simple yet distinctive appearance model for real time human tracking by exploiting the pairwise constraints between parts. The parts in our model are generated online by sampling the foreground of the scene into overlapping blocks and grouping them into appearance coherent parts with mean shift algorithm. Constraints between the resulting parts are defined and used to encode the structure of human body. To tolerate the possible human deformations and occlusions, the model is layered. With this model, we design an algorithm for human tracking and test its performance on real world image sequences. Experimental results show that the proposed appearance model although simple, has enough discriminative power to classify multiple humans even in presence of occlusions and the associated tracking method can run in real time.

Spatial color Mixture Of Gaussians model (SMOG model) based similarity measure is superior to the popular color histogram based one since it considers not only the colors in a region, but also the spatial layout of these colors. However, two drawbacks of SMOG are still obvious, firstly, in the initialization of SMOG, some background pixels are inevitably introduced and clustered as an object mode for tracking, this often degenerates the tracking performance. Secondly, the weight of each Gaussian mode is restricted by the probability of the pixels belong to it, so a low probability Gaussian mode always contribute a little in similarity measure even it has a high discrimination for discriminating the object. A revised SMOG model is proposed to efficiently cope with these two problems by sufficiently considering the object local background. Experiment results on synthetic and real image sequences verified the validity of the revised model.

Being put forward by the researchers in computer vision, self calibration commonly deals with camera with linear model. Since the distortion is practically existed especially for ordinary camera, the result of calibration can't meet the demand of vision measurement with high accuracy regardless of the distortion. Being obedience to systematism mainly, the distortion is the target function of distortion coefficient, principal point, principal distance ratio and skew factor etc. So there exists a group of parameters including of distortion coefficient, principal point, principal distance ratio and skew factor and fundamental matrix which make homologous point meets epipolar restriction theoretically. Accordingly, the paper advances the way titled self calibration of camera with non-linear imaging model which is on basis of the Kruppa equation. In calculating the fundamental matrix, we can obtain interior elements except principal distance by taking into account distortion correction about image coordinate. Then the principal distance can be obtained by using Kruppa equation. This way only need some homologous points between two images, not need any known information about objects. Lots of experiments have proven its correctness and reliability.

UAV Video is rapidly emerging as a widely used source of imagery for many applications in recent years. This paper presents our research on the UAV video processing system for the purpose of fire surveillance, which include: (1) UAV video stream processing. This step involves three aspects: decoding, re-sampling and matching. Microsoft(R) DirectX(R) technology is used to decode highly compressed video stream and re-sampled them into still video frame based on the time base and rate of UAV navigation sensor. One feature-based image-matching algorithm is developed to quickly obtain Tie points for latter calibration operation. (2) UAV system orientation. This step also involves three aspects: Camera IOP, Boresight Alignment and bundle adjustment. TSAI's two-stage technique is used to obtain initial camera focus length f, distortion coefficients k₁ and six Exterior Of Parameter (EOP) for one selected video image. Meanwhile, the Boresight Matrix is deduced by the comparison of GPS/INS derived parameters with solved EOPs. Further more, all parameters including EOPs of all re-sampled video images and camera IOP are optimally estimated based on developed bundle adjustment algorithm. (3) UAV Video geo-registration and mosaic. All re-sampled video frames are geo-registered into uniform geo-reference coordinate frame vice Classic photogrammetric orthorectification model and merge with each other with developed mosaic algorithm. The results demonstrated that the geo-accuracy of mosaic image generated from UAV video can achieve 1-2 pixels in planimetry and its combination with GIS-supported data for fast response to time-critical event, e.g., forest fire, is descried.

Wide baseline stereo correspondence has become a challenging and attractive problem in computer vision and its related applications. Getting high correct ratio initial matches is a very important step of general wide baseline stereo correspondence algorithm. Ferrari et al. suggested a voting scheme called topological filter in [3] to discard mismatches from initial matches, but they didn't give theoretical analysis of their method. Furthermore, the parameter of their scheme was uncertain. In this paper, we improved Ferraris' method based on our theoretical analysis, and presented a novel scheme called topologically clustering to discard mismatches. The proposed method has been tested using many famous wide baseline image pairs and the experimental results showed that the developed method can efficiently extract high correct ratio matches from low correct ratio initial matches for wide baseline image pairs.

An extended Bayesian classifier, which is able to fuse information in original image and in its wavelet domain, is designed for infrared image segmentation. The algorithm begins with a re-sampling process over the original image and a wavelet transformation of the original image. Then, the Spatially Variant Mixture Model (SVMM) is applied in the bootstrap samples and the wavelet coefficients. The corresponding parameters are estimated by EM (Expectation Maximum) algorithm. Finally, a two-element Bayesian classifier is constructed. One part of the classifier is designed to exploit information in the original image, and the other part is designed to exploit information obtained in the wavelet domain. Theoretic analysis and experimental results confirms that the approach is efficient for infrared image segmentation, robust to noise and less computationally involved.

The most methods of close-range photogrammetry are based on Direct Linear Transformation (DLT). But DLT often has unstable solution and every image needs more than six ground control points to compute DLT parameters, so this method is hard to acquire the high accuracy and its efficiency is low. The paper discusses a new method of digital close-range photogrammetry - panning and multi-baseline digital close-range photogrammetry. This method enlarges the intersection angle and improves the intersection precision by multi-baseline. At the same time this method applies the classic aerotriangulation and bundle adjustment to the close-range photogrammetry, we need more than three ground control points to compute the exterior orientation elements of all images. The experiments prove that this method can acquire the high accuracy.

A clustering method based on Joint Boost for Synthesis Aperture Radar images is proposed. In this method, we follow the steps of Joint Boost, but substitute weak learns with basic clustering algorithm. We compute the sharing features between samples in order to reduce clustering times. The proposed clustering method, JBC constructs a new training set by random sampling from the original dataset, then selects the best feature and the best clusters for sharing, and calculates a distribution over the training samples using current shared feature and clusters, and finally a basic clustering algorithm (e.g. K-mean) is applied to partition the new training set. The final clustering solution is produced by aggregating the obtained partitions. The clustering results for SAR images show that the proposed method has a good performance.

This paper presents a new distance measure for image matching based on local Kullback-Leibler divergence, which we call Image Kullback-Leibler Distance (IKLD). Unlike traditional methods, IKLD takes account into not only the spatial relationships of pixels, but also the structure information around pixels. Therefore, it is robust enough to small changes in viewpoint. In order to illustrate its performance, we imbed it into support vector machines for view-based object recognition. Experimental results based on the COIL-100 show that it outperforms most existing techniques, such as traditional PCA+LDA (principal component analysis, linear discriminant analysis), non-linear SVM, Discriminant Tensor Rank-One Decomposition (DTROD) and Sparse Network of Winnows (SNoW).

It is a challenging work to classify video shots into a predefined genre set according to their semantic contents, which is helpful to video indexing, summarization and retrieval. This research proposes a novel shot classification algorithm with concept detection for news video programs. Six semantic shot types are studied and categorized: Anchorperson, Monologue, Reporter, Commercial, Still image and Miscellaneous, in which anchorperson shots are detected by clustering methods, reporter and monologue shots are distinguished by Conditional Random Fields (CRFs), and the last three categories are picked out by rule-based methods. Multimodality features are employed, such as visual, audio, face, temporal and contextual features. The experimental results show its effectiveness and achieve a high average accuracy of 96.5%.

Scientists on the ground need understand the environment around the unmanned lunar rover in lunar exploration through analyzing data obtained by various payloads. There are two main material on the moon, high land material and mare material on the moon. We use reflectance spectrums of lunar soils from Apollo mission measured by LSCC to classify the two kinds of materials. Principal component analysis is applied to reduce and select the feature of the reflectance spectrums. These features input support vector machine, which base on statistical learning theory and is used widely to classify in modern pattern recognition. Our work shows that the reflectance spectrums of lunar soils are strong link with the material which they represent.

An algorithm for merging images has been proposed based on statistic principle of region geometric shape in the paper. The Algorithm is of high precision and speed for solving images conjoint when there is rare point's feature on those images. For enhancing the region geometric shape features we described an Image Difference Dynamic Binary Method firstly. And then the merging principle of images on which there is little point's feature. Last the process and steps of merging image are described in all details.

This paper deals with image quality analysis considering the impact of psychological factors involved in assessment. The attributes of image quality requirement were partitioned according to the visual perception characteristics and the preference of image quality were obtained by the factor analysis method. The features of image quality which support the subjective preference were identified, The adequacy of image is evidenced to be the top requirement issues to the display image quality improvement. The approach will be beneficial to the research of the image quality subjective quantitative assessment method.

Ant Colony Optimization (ACO) algorithm takes inspiration from the coordinated behavior of ant swarms, which has been applied in many study fields as a novel evolutionary technology to solve optimization problems. But it has rarely been used to process remote sensing data. Using the ACO algorithm to remote sensing image classification does not assume an underlying statistical distribution for the pixel data, the contextual information can be taken into account, and it has strong robustness. In this paper, taking Landsat TM data as an example, the process of ACO method in remote sensing data classification is introduced in detail, and has achieved a good result. The study results suggest that ACO become a new effective method for remote sensing data processing.

In the marker based human motion capture system, it's a key step to accurately extract and track the 2-D coordinates of the body joints, that because the 3-D reconstruction process and the reliability of the capture system depend heavily on it. Different from those traditional solutions, we use ordinary industrial cameras and take colorful balls as the markers to solve this key point. We have also promoted our solution to solve the problem of occlusion. Finally, we got perfect result in practical applications, and the whole process can be computed in real-time. The method will be extended in the future use.

A model of a measurement system composed by two CCD cameras using parallel binocular line-structured light is proposed. The singular value decomposition is used to solve the over-determined equation to obtain the parameters in the imaging model of the cameras. The feature recognition technique is applied to segment feature information of the image in the process of range image acquisition. Then pre-processing (image smoothing, binaryzation and image segmentation) of the image is processed, and the image is condensed to remove useless information. The image acquisition and condensing are carried out in parallel to gather image and extract effective data simultaneously. The proposed method solves the difficulty of removing the disturbed information in range image and realizes parallel data processing, which greatly simplifies the following work of image matching and image characteristic data extraction.

Confidence evaluation is an important technique in image matching process. This paper proposes a confidence level evaluation method for image matching result based on support vector machine (SVM). We divide the matching result into two different types: the correct result and the wrong result. So we translate the match result's confidence evaluation problem into the matching result's classification. This paper firstly provides a method of how to prepare the character parameters which can accurately reflect the matching performance. And then the SVM based on Gaussian kernel is used as a classifier to classify the match result and discriminate the match result's type. The experiments show that this method is effective. Compared with the Dempster-Shafer (D-S) evidence reasoning fusion method it has much higher accuracy.

Spectropolarimetric imaging can provide useful discriminating information for human face recognition that cannot be obtained by other imaging methods. This paper examines the ability of face recognition by using spectropolarimetric images. The Spectropolarimetric images were collected by using a CCD camera equipped with a liquid crystal tunable filter, which could capture 32 bands of images over the visible and near-infrared light (0.4μm-0.72μm). Since polarization techniques have better contrast mechanisms for tissue imaging and spectroscopy, and can also provide additional information about the structure of tissues, it is expected that better discriminate performance can be obtained by using polarimetric and spectral information than just using spectral information. An algorithm for facial characteristics analysis is presented to exploit only the spectropolarimetric information from different types of facial tissues. Experiments demonstrate that the proposed algorithm can distinguish efficiently the different facial tissues.

Nowadays, local feature based image categorization algorithm has attracted increasing attention in the computer vision community. In this paper, we present a local feature based image categorization scheme by using Multi-Scale Vocabulary. This technique works by partitioning the feature space into clusters at several different levels to form multi-scale vocabulary and generate corresponding fixed-length descriptors at different scales for each image. Then we design particular similarity measure for multi-scale descriptors and finally apply KNN and SVM to realize image categorization task. Experiments conducted on the ETH80 dataset have demonstrated the effectiveness of our approach.

Automatic target recognition(ATR) is the key of the image guidance technology, yet it is difficult to recognize the target by merely depending on the real-time image acquired by flying vehicle cameras, moreover, the task of recognizing the target from the real-time images by the vehicle-carrying image processing system is a hard work itself. The main trend of the ATR nowadays is to make utilization of the images produced by high-resolution remote sensing satellite to retrieve the front elevation of the interested region before hand. These front elevations are loaded upon the flying vehicles and are matched with the real-time images acquired by vehicle-carrying cameras to recognize the interested target. Obviously, the key step of this method is to recover the 3D information from 2D images. This paper proposed a framework to produce multi-scale and multi-viewpoint projection images based on remote sensing satellite stereopair by means of photogrammetry and computer vision. First we proposed a algorithm for reconstructing the 3D structure of the target by digital photogrammetric techniques and establishing the 3D model of the target using the OpenGL visualization toolkit. Then the conversion relationship between the world coordinate system and the simulation space coordinate system is provided to produce the front elevation in the simulation space.

In recent years, the tasks of fingerprint examiners have been greatly aided by the development of automatic fingerprint classification systems. These systems operate by matching low-level features automatically extracted from fingerprint images, often represented collectively as numeric vectors, for their decision. However, there are two major shortcomings in current systems. First, the result of classification depends solely on the chosen features and the algorithm that matches them. Second, the systems cannot adapt their results over time through interaction with individual fingerprint examiners who often have different degrees of experiences. In this paper, we demonstrate by incorporating relevance feedback in a fingerprint classification system, a personalized semantic space over the database of fingerprints for each user can be incrementally learned. The fingerprint features that induce the initial features space from which individual semantic spaces are being learned were obtained by multispectral decomposition of fingerprints using a bank of Gabor filters. In this learning framework, the out-of-sample extension of a recently introduced dimensionality reduction method, called Twin Kernel Embedding (TKE), is applied to learn both the semantic space and a mapping function for classifying novel fingerprints. Experimental results confirm this learning framework for examiner-centric fingerprint classification.

Automatic generate 3D models of buildings and other man-made structures from images has become a topic of increasing importance, those models may be in applications such as virtual reality, entertainment industry and urban planning. In this paper we address the main problems and available solution for the generation of 3D models from terrestrial images. We first generate a coarse planar model of the principal scene planes and then reconstruct windows to refine the building models. There are several points of novelty: first we reconstruct the coarse wire frame model use the line segments matching with epipolar geometry constraint; Secondly, we detect the position of all windows in the image and reconstruct the windows by established corner points correspondences between images, then add the windows to the coarse model to refine the building models. The strategy is illustrated on image triple of college building.

The development of the remote sensing technology makes us obtain very abundant information of nature, especially with the appearance of high resolution remote sensing image it extends the visual field of the nature. High-resolution satellite images such as Quickbird and IKONOS have been applied into many fields. But the challenge that faces us is how to make use of the data effectively and obtain more useful information through some processing. Because in the target recognition, the mutual-complementarity among the different results obtained by the different classifier making using of the same features usually is very strong and high resolution remote sensing data have a lot of characteristics such as spectral, texture and context and so on compared to the other lower resolution remote sensing data, the Multiple Classifiers making use of multi-characteristic was proposed to improve the high resolution remote sensing image classification accuracy in this paper. The experiments show that the approach can obtain higher classification accuracy and better classification result than single classifier.

With the aerial LIDAR technology developing, how to automatically recognize and reconstruct the buildings from LIDAR dataset is an important research topic along with the widespread applications of LIDAR data in city modeling, urban planning, etc.. Applying the information of the first-and-last echo data of the same laser point, in this paper, a scheme of 3D-reconstruction of simple building has been presented, which mainly include the following steps: the recognition of non-boundary building points and boundary building points and the generation of each building-point-cluster; the localization of the boundary of each building; the detection of the planes included in each cluster and the reconstruction of building in 3D form. Through experiment, it can be proved that for the LIDAR data with first-and-last echo information the scheme can effectively and efficiently 3D-reconstruct simple buildings, such as flat and gabled buildings.

In this paper, a face recognition method using local qualitative representations is proposed to solve the problem of face recognition in varying lighting. Based on the observation that the ordinal relationship between the average brightness of image regions pair is invariant under lighting changes, Local Binary Mapping is defined as an illumination invariant for face recognition based on Local Binary Pattern descriptor, which extracts the local variance features of an image. For the 'symbol' feature vector, hamming distance is used as similarity measurement. It has been proved that the proposed method can provide the accuracy of 100 percent for subset 2, 3, 4 and 98.89 percent for subset 5 of the Yale facial database B when all images in subset 1 are used as gallery.

The problem of object category recognition has long challenged the computer vision community. In this paper, we address these tasks via learning two-class and multi-class discriminative models. The proposed approach integrates the Adaboost algorithm into the decision tree structure, called DB-Tree, and each tree node combines a number of weak classifiers into a strong classifier (a conditional posterior probability). In the learning stage, each boosted classifier in a tree node is trained to split the training set to left and right sub-trees, and the classifier is thus used not to return the class of the sample but rather to assign the sample to the left or right sub-tree. Therefore, the DB-Tree can be built up automatically and recursively. In the testing stage, the posterior probability of each node is computed by the weighted conditional probability of left and right sub-trees. Thus, the top node of the tree can output the overall posterior probability. In addition, the multi-class and two-class learning procedures become unified, through treating the multi-class classification problem as a special two-class classification problem, and either a positive or negative label is assigned to each class in minimizing the total entropy in each node.

With the proliferation of multimedia information on the network, automatic rating of web pages becomes more and more important for web management. Obviously image analysis is very important in these kinds of tasks. But to our best knowledge, there are no publications reported on it. In this paper, we propose a novel framework to rate webpage using image content analysis. The rated categories are in compliance with the standard utilized by most web browsers like Internet Explorer, which include Normal, Revealing Attire, Exposed Breasts and Bare Buttocks. To make the rating work feasible, we analyze the images mainly using skin detection and body region, and face detection is also used as guidance for the detection of skin and body region. After that, all the results of image content analysis are integrated to achieve content rating results for web pages. We tested our system on two data sets and demonstrated its effectiveness.

Planar mapped trajectories (curves composed of time-stamped points) matching differs to other planar curves correspondence in stereo vision. Most methods proposed for general planar curves matching are not much suitable for it. In this paper, a new coarse-to-fine method for planar trajectories is proposed; some physical constraints such as time, height, speed limitation and continuousness of aerocraft trajectories are used to make matching process more reasonable. Similarity Measures between left and right images planar trajectories are calculated for matching, flight trajectories are reconstructed at last, and reconstruction error is also discussed at last.

In this paper an classification method based on polarimetric decomposition technique and neural network theory, is proposed for polarimetric SAR data sets. The main advantage of this polarimetric decomposition technique is to provide dominant polarimetric scattering properties identification information where the most important kinds of scattering medium can be discriminated. Feature vector extracted from full POLSAR data sets by polarimetric decomposition is used as input data of the feed-forward neural network (FNN). Neural networks have the advantage to be independent to the input signal statistics and the ability to combine many parameters in their inputs. To speed convergence and improve stability of the FNN Kalman filter plus scaled conjugate gradient algorithm is used in the training stage. The NASA/JPL AIRSAR c-band data of San Francisco is used to illustrate the effectiveness of the proposed approach to classification. Quantitative results of performance are provided, as compared to the Wishart classifier.

Fuzzy C-Means (FCM) algorithm is a fuzzy pattern recognition method. Clustering precision of the algorithm is affected by its equal partition trend for data set of large discrepancy of each class samples number, and the optimal clustering result of the algorithm mightn't be a right partition in this case. In order to overcome this disadvantage, a Gaussian function Weighted Fuzzy C-Means (WFCM) algorithm is proposed, which the weighted function is produced by a Gaussian function calculating dot density of each sample. To certain extent, the WFCM algorithm has not only overcome the limitation of equal partition trend in fuzzy Cmeans algorithm, but also been favorable convergence and stability. The calculation of the weighted function and the choice of sample dot density range restriction value for the algorithm are both objective. When partially supervised information obtained from a few labeled samples is introduced to the WFCM algorithm, the classification performance of the WFCM algorithm is further enhanced and the convergent speed of objective function is further accelerated.

This paper proposed a model for 2.1D sketch based on primalsketch [6,7] and a corresponding algorithm based on SWC(Swendsen-Wang Cut)[1,2]. Given a primal sketch graph, all boundaries' color and curvature information can be gotten from the dictionary of image primitives of the primal sketch. Suppose all curves in the sketch graph can be combined into closed contours and then contour can be combined into regions. Calculating boundaries similarity energy of two regions by comparing the comparability using the color and curvature information of boundaries, then the energy is regarded as the proposal of SWC. All regions are regarded as nodes of graph space, then the 2.1D sketch problem can be looked as graph partition problem. In each step, all curves will be reorganized into contours and then contours generate regions. Some experiment results based on above model and algorithm are given in the paper.

A novel improved broyden's method has been presented to estimate image jacobian matrix for uncalibrated visual servoing. In this paper, we apply chebyshev polynomial as a cost function to approximate best value. Compared with recursive least square (RLS) algorithm which is restricted by the prior knowledge for obtaining some performances, chebyshev polynomial algorithm has a great adaptability to estimate jacobian parameter, even without the prior knowledge. A microscopic image jacobian model has been developed for the four degree-of-freedom micromanipulator in our microassembly system. The performance has been confirmed by simulations and experiments.

Human chromosome karyotyping is one of the essential tasks in cytogenetics, especially in genetic syndrome diagnoses. In this thesis, an automatic procedure is introduced for human chromosome image analysis. According to different status of touching and overlapping chromosomes, several segmentation methods are proposed to achieve the best results. Medial axis is extracted by the middle point algorithm. Chromosome band is enhanced by the algorithm based on multiscale B-spline wavelets, extracted by average gray profile, gradient profile and shape profile, and calculated by the WDD (Weighted Density Distribution) descriptors. The multilayer classifier is used in classification. Experiment results demonstrate that the algorithms perform well.

This paper presents a line feature matching algorithm. Firstly, it extracts the set of line features in the image, and represents an object using attributed relational graph (ARG). By defining relation vectors between the adjacent features, the graph can describe the structural information of an object. Secondly, the one-to-one correspondences between model features and image features is established by two processes - coarse match and refined match through the analysis of matching ordering and matching number of relation vectors. Finally, the object examples in the image are extracted. Test showed that the proposed algorithm had superior performance to the present line feature matching algorithms, which is robust to shape deformation, or input noise, and decreased the computational cost.

The three-dimensional (3D) reconstruction has become a hot research field of Photogrammetry and computer vision. This paper presents a measurement method of CAD-based Photogrammetry to do 3D reconstruction. CAD models and images which the exterior orientation is known are used. The visible edges of a volumetric CAD model with an initial pose are projected back into images after a hidden line analysis. This analysis determines which edges (or parts of edges) of the object are visible, and computes the edges of the object. Observation equations are formulated that establish the relationship between the parameterized CAD models and position of the edges in the images. The flow of this paper is as fellows. Firstly, parameterized CAD models are used to describe 3D objects, object models can be described in various ways, and this paper uses CSG models combined with B-rep. Secondly, fitting models to images. Manually select an appropriate model and approximate parameters, the parameterized models are projected back into the images, and then compute the optimal fitting automatically. Thirdly, calculate the parameters of the model. Experiments were done to box, sphere and column, which are the base primitive of some objects. The results of these experiments show that the method used in this paper are effective and have high precision. All primitives used in these experiments are well reconstructed.

The inhibitory interaction has long been observed in the lateral eye of the Limulus and been integrated into mechanism of enhancing contrast. When applying to the enhancement of low-contrast image for segmenting interested objects, the original lateral inhibition model will simultaneously amplify noises while enhancing edges contrast. This paper presents a new lateral inhibition model, which is called Stick-Guided Lateral Inhibition, for enhancement of low-contrast image so that week edges may exert a stronger force to catch the boundary of targets in the latter segmentation. First, the guided inhibition term is introduced as a general framework for improving the performance of lateral inhibition models in the presence of noises. Then, by using asymmetric sticks to guide the inhibiting process, we are able to accentuate the intensity gradients of image-edges and contours while suppressing the amplification of noises. Experiments on synthetic images and remote sensor images show that our model significantly enhances low-contrast images and improves the performance of latter segmentation.

Reentry guidance is one of the key technologies in hypersonic vehicle research field. In addition to the constraints on its final position coordinates, the vehicle must also impact the target from a specified direction with high precision. And therefore the adaptability of guidance law is critical to control the velocity of hypersonic vehicle and firing accuracy properly in different surroundings of large airspace. In this paper, a new adaptive guidance strategy based on Back Propagation (BP) neural network for the reentry mission of a generic hypersonic vehicle is presented. Depending on the nicer self-learn ability of BP neural network, the guidance law considers the influence of biggish mis-modeling of aerodynamics, structure error and other initial disturbances on the flight capability of vehicle. Consequently, terminal position accuracy and velocity are guaranteed, while many constraints are satisfied. Numerical simulation results clearly bring out the fact that the proposed reentry guidance law based on BP neural network is rational and effective.

In recent years, there has been a growing interest in the research of image content description techniques. Among those, image clustering is one of the most frequently discussed topics. Similar to image recognition, image clustering is also a high-level representation technique. However it focuses on the coarse categorization rather than the accurate recognition. Based on wavelet transform (WT) and advanced statistics, the authors propose a novel approach that divides various shaped logo images into groups according to the external boundary of each logo image. Experimental results show that the presented method is accurate, fast and insensitive to defects.

Semiconductor visual inspection is necessary for production yield control. Defect classification is a key procedure in determing defect sources. Auttomization of this procedure is required in order to achieve efficient and high-yield production. In the present paper, an automatic defect classification (ADC) algorithm for a semiconductor inspection is proposed. The ADC algorithm consists of the following three parts; 1) A defect extraction algorithm to achieve high-sensitivity defect extraction even in regions in which the brightness is unstable due to optical interference at a thin layer. 2) An appearance feature calculation from a color image inside the defect region extracted from 1). 3) A unique training type classifier called the fuzzy selective voting classifier (FSVC), which calculates the weight for each appearance feature in order to achieve accurate classification even when the discriminancy of each feature is different. The performance of the developed ADC algorithm has been evaluated using defect acquired from an actual production line. The accuracy of the classification was 85.9% and the false rejection rate was 93%.

A novel algorithm for matching synthetic aperture radar (SAR) image to Optical image based on lineal feature using Hausdorff distance combined with genetic algorithm is proposed in this paper. A new method is presented to extract lineal feature from low signal to noise ratio (SNR) SAR image. Based on the edge image from SAR and Optical image, modified Hausdorff distance is adopted as a similarity measure because it is insensitive to noise. Genetic algorithm is used as searching strategy to achieve high computation speed for its inertial parallel. Experimental results using real SAR and Optical images demonstrate that the algorithm is robust, fast and can achieve high matching accuracy.

A new similarity measure based on Hausdorff Distance Matrix Frobenius Norm for object matching is proposed in this paper. This measure is more reliable and can achieve higher location accuracy compared with other measures based on classic and modified Hausdorff Distance under the condition of high level noise and high ratio occlusion of template. The search strategy based on genetic algorithms is employed to make algorithm faster. Experimental results under noise of different level demonstrate high performance of the matching algorithm.

A new nonlinear control strategy incorporated the differential geometry method with adaptive neural networks is presented for the nonlinear coupling system of Bank-to-Turn missile in reentry phase. The basic control law is designed using the differential geometry feedback linearization method, and the online learning neural networks are used to compensate the system errors due to aerodynamic parameter errors and external disturbance in view of the arbitrary nonlinear mapping and rapid online learning ability for multi-layer neural networks. The online weights and thresholds tuning rules are deduced according to the tracking error performance functions by Levenberg-Marquardt algorithm, which will make the learning process faster and more stable. The six degree of freedom simulation results show that the attitude angles can track the desired trajectory precisely. It means that the proposed strategy effectively enhance the stability, the tracking performance and the robustness of the control system.

Owing to its fast speed, simple operation, and strong robustness, Shape Distribution is widely used in search engines. This method, however, only considers distances between the objects' shape distribution histograms and ignores the information included. Actually the information of the shape distribution histograms, such as the mean value, the standard deviation, the kurtosis and the skewness, can be used to map the 3D model. As a result, the retrieval precision of Shape Distribution is low. To enhance the retrieve efficiency, a novel method which employs the K-means clustering method is proposed in this paper. First, the models' shape distribution histograms are established by Shape Distribution method and are normalized as the proper format of K-means clustering method. Then, the objects' shape distribution histograms are served as inputs of K-means clustering method and are classified into certain groups by this algorithm. Last, all the models that belong to the classification of the query model are exported as the retrieval results. A case study is used to validate the proposed method. Experimental results show that the retrieval precision by using the proposed method is higher than that of the Shape Distribution method.

The problem of identifying spectra collected by large sky survey telescope is urgent to study to help astronomers discover new celestial bodies. Due to spectral data characteristics of high-dimension and volume, principle component analysis (PCA) technique is commonly used for extracting features and saving operations. Like many other matrix factorization methods, PCA lacks intuitive meaning because of its negativity. In this paper, non-negative matrix factorization (NMF) technique distinguished from PCA by its use of nonnegative constrains is applied to stellar spectral type classification. Firstly, NMF was used to extract features and compress data. Then an efficient classifier based on distance metric was designed to identify stellar types using the compressed data. The experiment results show that the proposed method has good performance over more than 70,000 real stellar data of Sloan Digital Sky Survey (SDSS). And the method is promising for large sky survey telescope projects.

One of the key technologies of Outdoor AR is the real-time 3D registration of objects in the real world. The paper puts forward a new hardware registration method which not only borrows the ideas of identification point registration, but improves it to realize tracking registration in video-based outdoor AR, which uses see-through head mounted display (STHMD) loaded on outdoor AR system for showing the result of registration, and employs one color CCD camera capturing video to obtain the world coordinate of scene border. Furthermore, the paper utilizes 3D electronic compass and GPS attached on user's body to calculate transition matrix from the world coordinate system to the camera coordinate system. Then, the transition matrix from the virtual coordinate system to the image plane can be calculated out and 3D virtual object generated by computer model is added into the STHMD as a whole. Synthetically, video-based registration offers a superior approach to 3D registration of dynamic object. Finally, the paper provides the implementation process and designs a test. By the case study, the new method significantly simplifies the registration system and algorithm, and coordination errors are eliminated. The algorithm requires little computation and can be easily realized in real time without delay. Compared with the several existing registration methods, it is significantly improved.

Skin is a widely used cue for porn image classification. Most conventional methods are off-line training schemes. They usually use a fixed boundary to segment skin regions in the images and are effective only in restricted conditions: e.g. good lightness and unique human race. This paper presents an adaptive online training scheme for skin detection which can handle these tough cases. In our approach, skin detection is considered as a classification problem on Gaussian mixture model. For each image, human face is detected and the face color is used to establish a primary estimation of skin color distribution. Then an adaptive online training algorithm is used to find the real boundary between skin color and background color in current image. Experimental results on 450 images showed that the proposed method is more robust in general situations than the conventional ones.

In this paper an ensemble feature extraction algorithm is proposed based on Adaboost.M2 for multiclass classification problem. The proposed algorithm makes no assumption about the distribution of the data and primarily performs by directly selecting the discriminant features with the minimum training error, which can overcome the main drawbacks of the traditional methods, such as Principle Component Analysis (PCA), Fisher Linear Discriminant Analysis (FLD) and Nonparametric Discriminant Analysis (NDA). The proposed method first samples a large number of bootstrap training subsets from the original training set and implements FLD in each subset to get a large number of bootstrap FLD projections. Then at each step of Adaboost.M2, the projection with the minimum weighted K Nearest Neighbor (KNN) classification error is selected from a pool of linear projections to combine the final strong classifier. Experimental results on three real-world data sets demonstrate that the proposed algorithm is superior to other feature extraction techniques.

In this paper we present a novel self-organizing shape from shading method based on hybrid reflection model which includes diffuse and specular reflection components. Using this method, the shape of an object surface is recovered in two steps. Firstly, a grayscale image of the surface is separated into diffuse and specular components, and a new image only composed of diffuse component is created. Secondly, a self-organizing shape from shading algorithm is performed for the new image, and the 3D shape of the surface is generated. The experimental results for synthetic and real images demonstrate the availability and feasibility of our method.

Support Vector Data Description (SVDD) is a one-class classification method developed in recent years. It has been used in many fields because of its good performance and high executive efficiency when there are only one-class training samples. It has been proven that SVDD has less support vector numbers, less optimization time and faster testing speed than those of two-class classifier such as SVM. At present, researches and acquirable literatures about SVDD multi-class classification are little, which restricts the SVDD application. One SVDD multi-class classification algorithm is proposed in the paper. Based on minimum distance classification rule, the misclassification in multi-class classification is well solved and by applying the threshold strategy the rejection in multi-class classification is greatly alleviated. Finally, by classifying range profiles of three targets, the effect of kernel function parameter and SNR on the proposed algorithm is investigated and the effectiveness of the algorithm is testified by quantities of experiments.

Clustering is a fundamental problem for a great variety fields such as pattern recognition, computer vision. A popular technique for clustering is based on K-means. However, it suffers from the four main disadvantages. Firstly, it is slow and scales poorly on the time. Secondly, it is often impractical to expect a user to specify the number of clusters. Thirdly, it may find worse local optima. Lastly, its performance heavily depends on the initial clustering centers. To overcome the above four disadvantages simultaneously, an effectively adaptive K-means clustering algorithm (AKM) is proposed in this paper. The AKM estimates the correct number of clusters and obtains the initial centers by the segmentation of the norm histogram in the linear normed space consisting of the data set, and then performs the local improvement heuristic algorithm for K-means clustering in order to avoid the local optima. Moreover, the kd-tree is used to store the data set for improving the speed. The AKM was tested on the synthetic data sets and the real images. The experimental results demonstrate the AKM outperforms the existing methods.

In fuzzy clustering, we always use fuzzy equivalence matrix to find the clustering results. Although fuzzy equivalence matrix is one of the most important tools in fuzzy clustering, it's hard to judge whether a fuzzy matrix is a fuzzy equivalence matrix directly. In this paper, some simple and useful properties of the fuzzy equivalence matrices have been proposed based on the basic properties proposed in the classical articles. Using these properties, the judgment whether a fuzzy matrix is a fuzzy equivalence matrix may become an easy work.

Accuracy is one of the most important features for a 3D shape measurement system based on gray-code and phase-shift structured light projection and it requires accurate modeling and calibration of the camera and projector that consist of the system. In this paper, a new method is proposed to reconstruct the 3D object without the inverse projector model, thus make it possible to use a more accurate projector model considering high to 4th order radial and tangential lens distortion. Moreover, we propose a new system calibration method which can calibrate the camera and projector simultaneously based on the same reference points to eliminate the disadvantage of the conventional two-step calibration methods that the accuracy of the projector calibration is always influenced by the camera calibration error. The experiments compare the performance of our proposed model and calibration method with the conventional ones and the results are presented.

The dynamic range of many real-world environments exceeds the capabilities of current display technology by several orders of magnitude. To obtain reasonable reproduction, a large number of high-quality tone-mapping operators are currently available. All of these methods can be divided into three kinds: global operation, local operation and temporal-related algorithm. As an effective local operation to avoid halo artifacts, bi-lateral filter is presented and discussed in recent years. After analysis in-depth, this paper presents an improved bi-lateral filter in high dynamic range compression focused on four points: calculating efficiency, vision theory support, scales and parameters. Experiments indicate that the new operator can generate reasonable reproduction of high dynamic range images.

The aim of this study is to investigate the potential of the visible and near infrared spectroscopy (Vis/NIRS) technique for non-destructive measurement of soluble solids contents (SSC) in grape juice beverage. 380 samples were studied in this paper. Smoothing way of Savitzky-Golay and standard normal variate were applied for the pre-processing of spectral data. Least-squares support vector machines (LS-SVM) with RBF kernel function was applied to developing the SSC prediction model based on the Vis/NIRS absorbance data. The determination coefficient for prediction (Rp²) of the results predicted by LS-SVM model was 0. 962 and root mean square error (RMSEP) was 0. 434137. It is concluded that Vis/NIRS technique can quantify the SSC of grape juice beverage fast and non-destructively.. At the same time, LS-SVM model was compared with PLS and back propagation neural network (BP-NN) methods. The results showed that LS-SVM was superior to the conventional linear and non-linear methods in predicting SSC of grape juice beverage. In this study, the generation ability of LS-SVM, PLS and BP-NN models were also investigated. It is concluded that LS-SVM regression method is a promising technique for chemometrics in quantitative prediction.

Much attention is paid to registration of terrestrial point clouds nowadays. Research is carried out towards improved efficiency and automation of the registration process. The most important part of registration is finding correspondence. The panoramic reflectance images are generated according to the angular coordinates and reflectance value of each 3D point of 360° full scans. Since it is similar to a black and white photo, it is possible to implement image matching on this kind of images. Therefore, this paper reports a new corresponding point matching algorithm for panoramic reflectance images. Firstly SIFT (Scale Invariant Feature Transform) method is employed for extracting distinctive invariant features from panoramic images that can be used to perform reliable matching between different views of an object or scene. The correspondences are identified by finding the nearest neighbors of each keypoint form the first image among those in the second image afterwards. The rigid geometric invariance derived from point cloud is used to prune false correspondences. Finally, an iterative process is employed to include more new matches for transformation parameters computation until the computation accuracy reaches predefined accuracy threshold. The approach is tested with panoramic reflectance images (indoor and outdoor scenes) acquired by the laser scanner FARO LS 880. ¹

The classification criterion for the two dimensional LDA (2DLDA)-based face recognition methods has been little considered, while we almost pay all attention to the 2DLDA-based feature extraction. The typical classification measure used in 2DLDA-based face recognition is the sum of the Euclidean distance between two feature vectors in feature matrix, called traditional distance measure (TDM). However, this classification criterion does not match the high dimensional geometry space theory. So we apply the volume measure (VM), which is based on the high dimensional geometry theory, to the 2DLDA-based face recognition in this paper. To test its performance, experiments were performed on the YALE face databases. The experimental results show the volume measure (VM) is more efficient than the TDM in 2DLDA-based face recognition.

There are reported methods for fire or smoke recognition respectively. But in forest, fire and smoke always exist together, and fire area lies within smoke area. Both fire and smoke are important features for fire detection. Based on this fact we present a novel method to detect fire and smoke in two steps and obtain areas of fire and smoke together. With the help of Otsu method taking gray value and red value as inputs, fire and smoke regions are segmented from the background, and regions with very small areas are deleted as noises; then fire is segmented from the left large and continuous regions. Area, roundness and contour of segmented results are used to further recognize smoke or fire from other regular objects, or to describe the status and trend of forest fire.

This paper introduces an improved graph cut approach and combines it with commonly used surface texture representations for synthesis. Unlike graph cut which supports iterative refinement for improvement of the patch seams, the improved graph cut method re-synthesis the max error overlap region using k-nearest neighbor method. This paper follows a simple 3D surface texture synthesis framework, which comprises three steps: capture, synthesis and relighting. Firstly, it represents 3D surface textures under varied illumination directions using gradient and albedo maps. Then, it uses the surface representation to synthesis a description of a larger surface. Finally, it renders the surface representation. It proposes an improved graph cut method for the synthesis step, and the proposed method can generate perceptually smooth images of 3D surface texture from small samples.

A gradient-based face recognition method using Partial Hausdorff Distance (PHD) measure is proposed in this paper. First, in order to achieve a performance independent of lighting conditions, the image is transformed into a Gradient Map (GM). And then, Hausdorff distance measure is introduced to calculate the dissimilarity between two Gradient Maps. The experimental data show that the measure is suitable for face recognition. As we can see later, this distance measure is robust to lighting variations, slight pose differences and expression changes in face images. At last, recognition accuracy is given tested on AR and FERET databases, and comparisons with Edge Map (EM) and Line segment Edge Map (LEM) approaches are also presented.

With the development of non-contact measurement in the close-range photogrammetry, the use of the projector becomes frequent and familiar. In order to take full advantages of the slide projector, the slide projector is steered on the basis of the traditional method of the digital camera taking images. The paper proposes a flexible and effective technique for the slide projector calibration based on the calibration of digital camera. The technique only requires a planar grid, a rotating platform and a computer. The planar grid is put on the center of the rotating platform functioned as a control ground. The computer can control the rotating platform to rotate regularly and dominate the slide projector and the digital camera to project slides and take images automatically and synchronously. The algorithm with 2D direct linear transformation (2D-DLT) and collinear equations is used to calibrate the digital camera first. By the similar algorithm above, the interior parameters of the slide projector can be determined too. Then the calibration of the slide projector is achieved finally. The operation method in detail and the algorithm are addressed systematically and entirely. The feasibility and the exactness of the technique for the calibration of the slide projector based on the digital camera calibration put forward in this paper are verified by the results of real experimental data.

Camera self-calibration is one key problem in 3D reconstruction from image sequences. In this paper, a further discussion is given for self-calibration method of Absolute Dual Quadric (ADQ). This method using projective matrix as known data to extract camera intrinsic parameters, however, projective matrix still contains some uncertain variables, and optimization to them were ignored, in addition, the utilities of constraint conditions are not convincible in nonlinear algorithm of this method. An improved algorithm is given to resolve these problems in this paper. The main idea of the improved algorithm is to recalculate projective matrix with middle calibration result in each iterative process of the algorithm, and set convergence of iterative process to the point on parabolas where the ratio of initial focal length to its result reach to 1.0. Experiment result from real image data shows that improved algorithm is efficient.

When using ontological theory to set up remote sensing image knowledge system, the majority of scholars by now regard ontology as logical theory for defining the object, attribution relation, affair and process of remote sending knowledge system. But understanding and describing the real world makes that logical theory be unable to unify concepts with the same practical meaning from different concept models, so bring grid service system drawbacks in knowledge delivering and sharing. To solve that issue requires further improvement of the defining method and model for concepts in ontology. This paper presents a neural network remote sensing image ontological concept extraction model based on image understanding and describing, utilize the theory of bionic optimization, and adopts the combination of artificial neural network with the rule-based knowledge Recognition System. Realize the knowledge delivering and sharing among different information systems or make the knowledge delivering and sharing between client and system possible and effective.

Generally, while designing pattern classifier, the boundaries between different classes are vague and it is often difficult or impossible to acquire all of the necessary essential features for precisely classifying, so often both the fuzzy uncertainty and rough uncertainty are exist in classification problems. In this work, a novel FRMFN (Fuzzy-Rough Membership Function Neural Network) is built based on fuzzy-rough sets theory. The FRMFN integrates the ability of processing fuzzy and rough information simultaneously. The test results of classification for infrared band combination image of Canada Norman Wells area and five vowel characters indicate that FRMFN has better classification precision than RBFN (Radial Basis Function Neural Network) and has the same merit of quick learning as RBFN.

Hyperspectral RS technology organically combines the radiation information and the space information. The spectrum information, which the hyperspectral image enriches, can be better to carry on the ground target classification, compare with panchromatic remote sensing image and multispectral remote sensing image. As support vector machine was applied to many fields successfully recent years, using kernel methods, the classic linear methods can cope with the nonlinear problem, which was called the 3rd revolution of pattern analysis algorithms. This paper introduced two classifying methods for hyperspectral image based on kernel function, Support Vector Machine and Kernel Fisher Discriminant Analysis, and studied the selection of kernel function and its parameters as well as multi-class decomposition. We use radial basic function kernel, one against one or one against rest decomposition methods to construct multi-class classifier, and optimize parameter selection using cross-validating grid search to build an effective and robust kernel classifier. It is verified that, through the OMIS and AVIRIS image classifying experiments, comparing with common image classifying methods, kernel classifying method can avoid Hughes phenomenon, thus improve the classifying accuracy.

A novel approach combining 2DCCA, edge detector, and corner detector for object detection is proposed in this paper. The detection system consists of two stages. In the first stage, edge and corner information is obtained by edge detector and corner detector. By setting range for the number of edge pixel and corner in the scanning window, a large number of non-object windows are rejected. In the second stage, the classifier trained by 2DCCA is combined with slide window method so that further non-object windows are rejected. For the case that one object is simultaneously contained in several windows, the algorithm of determining the best position of object is designed. Compared with related approaches, our method has advantage of obtaining higher precision under the similar recall. The performance of the proposed approach is illustrated by experimental results.

We outline a platform for rendering the canopy reflected irradiance under the irradiation of different wavelengths. Our approach is to use different colors to symbol value of canopy reflected irradiance in different ranges. Meanwhile, we do some statistical analysis according to the results above. We illustrate our approach for simple 3-D trees scenes under the irradiation of 6 different wavelengths, respectively in visible light waveband and near infrared waveband. Our results are good. We get the vivid light shadow effect in the representation of 3-D trees scenes. We use 6 colors to represent 6 consecutive canopy reflected irradiance range and make some statistical analysis to get useful information from the results above.

During the process of watermelon seeds characteristic extraction as well as separation, watermelon seeds' major and minor axes, the length and width ratio have played a very important role in appearance regulating degree evaluation. It is quite difficult to find the answer of orthogonal bi-axes because the watermelon seeds are flat and irregular in shape and what's more there is no rule to follow. After a lot of experiments and research, the author proposed the algorithm of orthogonal bi-axes algorithm for granulated object. It has been put into practice and proved in the application of auto-separation system for watermelon seeds. This algorithm has the advantage of lower time complexity and higher precision compared with other algorithms. The algorithm can be used in the solution of other similar granulated objects, and has the widespread application value.

The study documented successful discrimination between five weed species and rapeseed plants under actual field conditions using visible and near infrared (Vis/NIR) spectroscopy. A hybrid recognition model, BP artificial neural networks (BP-ANN) combined with principal component analysis (PCA), had been established for discrimination of weeds in rapeseed field. Spectra tests were performed on the rapeseed and five-weed species canopy of 180 samples in the field using a spectrophotometer (325-1075 nm). 6 optimal PCs were selected as the input of BP neural networks to build the prediction model. Rapeseed samples were marked as 1, while the five weed species marked as 2, 3, 4, 5, 6, which were used as output set of BP-ANN. 120 samples were randomly selected as the training set, and the remainder as prediction set. It showed excellent predictions with the correlation value of 0.9745, and the relative standard deviation (RSD) was under 5% thus 100% of prediction accuracy was achieved. The results are promising for further work in real-time identification of weed patches in rapeseed fields for precision weed management.

Near-infrared spectroscopy (NIRS) with the characteristics of high speed, non-destructiveness, high precision and reliable detection data, etc. is a pollution-free, rapid, quantitative and qualitative analysis method. A new approach for variety discrimination of brown sugars using short-wave NIR spectroscopy (800-1050nm) was developed in this work. The relationship between the absorbance spectra and brown sugar varieties was established. The spectral data were compressed by the principal component analysis (PCA). The resulting features can be visualized in principal component (PC) space, which can lead to discovery of structures correlative with the different class of spectral samples. It appears to provide a reasonable variety clustering of brown sugars. The 2-D PCs plot obtained using the first two PCs can be used for the pattern recognition. Least-squares support vector machines (LS-SVM) was applied to solve the multivariate calibration problems in a relatively fast way. The work has shown that short-wave NIR spectroscopy technique is available for the brand identification of brown sugar, and LS-SVM has the better identification ability than PLS when the calibration set is small.

In order to implement the fast discrimination of different milk tea powders with different internal qualities, visible and near infrared (Vis/NIR) spectroscopy combined with effective wavelengths (EWs) and BP neural network (BPNN) was investigated as a new approach. Five brands of milk teas were obtained and 225 samples were selected randomly for the calibration set, while 75 samples for the validation set. The EWs were selected according to x-loading weights and regression coefficients by PLS analysis after some preprocessing. A total of 18 EWs (400, 401, 452, 453, 502, 503, 534, 535, 594, 595, 635, 636, 688, 689, 987, 988, 995 and 996 nm) were selected as the inputs of BPNN model. The performance was validated by the calibration and validation sets. The threshold error of prediction was set as ±0.1 and an excellent precision and recognition ratio of 100% for calibration set and 98.7% for validation set were achieved. The prediction results indicated that the EWs reflected the main characteristics of milk tea of different brands based on Vis/NIR spectroscopy and BPNN model, and the EWs would be useful for the development of portable instrument to discriminate the variety and detect the adulteration of instant milk tea powders.

To help solve the problems existed in current RP systems using STL data model, which are low precision, hard to distribute efficiently among various CAX systems, and to establish some fundament for RP system based on STEP-NIC, this article studied the gathering method of 3-d geometric model feature of STEP-NC. Based on the in-depth analysis of AP203 neutral file's structure and its description language EXPRESS, a data gathering module for AP203 neutral exchange file was made in the form of a dynamic link library, achieved the goal of correctly acquiring the STEP-NC 3-d CAD model's geometric and topologic feature information. This provides the fundament for achieving slicing of STEP-NC and developing corresponding RP system.