This PDF file contains the front matter associated with SPIE Proceedings Volume 8004, including the Title Page, Copyright information, Table of Contents, and the Conference Committee listing and Introduction.

Transfer estimation is a new parameter estimation method, which utilizes samples from not only the target distribution but also related ones. It is based on traditional estimators and can be used to solve the problem of small sample estimation. The key problem in transfer estimation is how to set the weighting coefficients when designing the transfer estimators. In this paper, we will propose a new algorithm to solve this problem. First, we will propose a common rule. Following the proposed rule, we can formulate the weighting coefficient setting problem as a constrained optimization problem. We will introduce an alternating optimized method and get a new algorithm of transfer estimation. The estimation of evolving class priors in data stream classification is a typical and important small sample estimation problem. In this paper, we will apply the proposed transfer estimation algorithm to the class prior estimation. Experiments on benchmark data sets will be performed, which show that the proposed algorithm can improve the performance on both class prior estimation and the final data stream classification.

A novel randomized clustering method is proposed to overcome some of the drawbacks of Mean Shift method. A hypothetical potential field is constructed from all the data points. Different from Mean Shift which moves the kernel window towards high-density region, our method moves the kernel window towards low-potential region. The proposed method is evaluated by comparing with both Mean Shift and K-means++ on three synthetic data sets which represent the clusters of different sizes, different shapes and different distributions. The experiments show that our method can produce more accurate results than both Mean Shift and K-means++.

In the linear discriminative analysis, especially in the high dimension case, it is insufficient to project the data onto a one-dimensional subspace for the two-category classification problem. Therefor a weak component approach (WCA) was proposed to project patterns to a low dimensional subspace with rich number of classification features. The role of the weak component in pattern classification was discussed. And the abundance of discriminative information contained in weak components was explored. Firstly, a definition of the weak component was given. Secondly, an improved regularization method was proposed. The regularization is a biased estimate of the variance in the corresponding dimension of the training data and the population data. Then a construction method of the feature subspace based on weak component was given, which extracts the eigenvector of the scatter matrixes according to their discriminative information. Finally, the proposed approach was validated in the experiments by comparing it with LDA. A better classification accuracy of the presented method was achieved. As WCA extracts the dims on which the data distributes closer, it is applicable to the high-dimensional data which distributes elliptically.

Fingerprint matching is the most important part in the field of fingerprint recognition. In this paper, a novel fingerprint matching algorithm based on the probabilistic graphical model and 3-tree model is proposed. First, minutiae matching problems are considered as a special point-set matching. Fingerprint minutiae are viewed as random variables. Each minutia pairs have some probability to be matched. Second, an algorithm is proposed to generate the graphical model and choose "signal points", which dynamically have corresponding points in other point set. We choose three base minutiae pairs as signal pairs. Third, the model is converted into a Junction Tree. A 3-tree model is built and the potentials of other minutiae pairs are calculated through Junction Tree (J.T.) algorithm. Then we translate the matching problem into the best matching problem of a weighted bipartite graph. Finally, the number of common matching pairs can be got through maximum flow algorithm. The similarity of two fingerprints is evaluated using the number of common matching pairs and the maximal posteriori probability. In order to deal with part-matching problems, we use the smallest convex hull which contains all the matched minutiae. Experiments evaluated on FVC 2004 show both effectiveness and efficiency of our methods.

Cut is a common type of shot boundary. In previous literature, frame pair similarity (FPS) was usually used as building block of cut detector. For a given frame, one needs to determine what frame pairs in its adjacency to select and how to combine the FPS values. To do these, previous methods, including Cross Analysis, Full Analysis, Graph Cut, etc, resort to human experience and thus lead to inferior performance. Besides, they are susceptible to noise caused by flashlight or popup subtitle such that additional procedure is needed to suppress such noise. In this paper we propose a novel framework to address these problems. Both the frame pair selection and the similarity values combination are done via machine learning. Features insensitive to flashlight or popup subtitle are extracted by exploiting the color histogram based FPS. Experimental results on TRECVID2003~2005 datasets demonstrate the effectiveness of the proposed algorithm.

This paper studied an algorithm on image matching based on Chang'e-1 CCD image, the model used the three-line-array image to do image matching and the constraint condition of the quasi-epipolar line on the course of matching. The experiment proved the model is very availably and can get better effect.

Currently, social tagging systems have been adopted by many social websites. As tags help users to browse social content effectively, personalized tag prediction problem becomes important in social networks. In this paper, we present a new generative probabilistic model to solve personalized tag prediction problem. Differently with previous methods, we consider social influence between users and friends into this model. We bring two major contributions: 1) We propose a new probabilistic model which considers in social influence to describe users' actual tagging activities; 2) Based on this model, we propose a new approach to perform personalized tag prediction task. Experimental results on a real-world dataset crawled from Last.fm show that our method outperforms other methods.

Tracking objects in videos using mean shift technique has brought to public attention. In this paper, we developed an improved tracking algorithm based on the mean shift framework. To represent the object model more accurately, the motion direction of the object which was estimated by the local motion filters was employed to weight the histogram. Besides, a wise object template updating strategy was proposed to adapt to the change of the object appearance caused by noise, deformation or occlusion. The experimental results on several real world scenarios shows that our approach has an excellent tracking performance comparing with the background weighted histogram mean shift tracking approach and traditional mean shift tracking method.

Accurate modeling of sea clutter and robust detection of low observable targets within sea clutter are the important problems in the field of remote sensing and radar signal processing. In this paper, a multiplicative multifractal process is introduced for the modeling sea clutter amplitude time series; a double parameter model which outperforms the double exponential model and Gaussian model is proposed and applicable for distribution fitting for multiplier of multiplicative multifractal modeling of real sea clutter. In addition, a short time generalized dimension spectrum based measurements of matching goodness is proposed and analyzed carefully, new model for HF sea clutter is built and reasonability of STGDs based measurements has been proven by simulation analysis. This modeling is computable rapidly and applicable in the research of target detection in HF sea-clutter background.

This paper presents a different learning-based image super-resolution enhancement method based on blind sparse decomposition, in order to improve its resolution of a degraded one. Firstly, sparse decomposition based image super-resolution enhancement model is put forward according to the geometrical invariability of local image structures under different conditions of resolution. Secondly, for reducing the complexity of dictionary learning and enhancing adaptive representation ability of dictionary atoms, the over-complete dictionary is constructed using online learning fashion of the given low resolution image. Thirdly, since the fixed sparsity of the conventional matching pursuit algorithms for sparse decomposition can not fit all types of patches, the approach to sparse decomposition with blind sparsity can achieve relatively higher accurate sparse representation of an image patch. Lastly, atoms of high resolution dictionary and coefficients of representation of the given low-resolution image are synthesized to the desired SR image. Experimental results of the synthetic and real data demonstrate that the suggested framework can eliminate blurring degradation and annoying edge artifacts in the resulting images. The proposed method can be effectively applied to resolution enhancement of the single-frame low-resolution image.

This paper presents a novel fast template matching algorithm based on context prediction. The predicted regions are those windows that contain the current entire sub-window. Comparison skipping or comparison terminating is executed when a low bound of distance which has been calculated between the template and the window exceeds the threshold. Experimental results and theory analyses prove the proposed method is faster than the conventional fast template matching method, strictly guaranteeing the same accuracy and up to maximal twenty times faster than the SSDA.

A reference image creating approach is proposed in this paper for contour-based infrared ATR algorithm. The approach based on three-dimensional realistic scene modeling consists of five parts: three-dimensional scene model construction, view plane simulation, coordinates transformation and perspective projection, visible-surface identification and three-dimensional rendering algorithm, contour extraction algorithm. By using this approach, a distinctively reference image is rapidly generated at multiple objective views and different scales. Numerical experiments are chosen to test proposed method from velocity performance. The experimental results indicate that the performance of the proposed technique is better than that of other methods.

Image matching is one of the key technologies in image processing applications,and the problem about the matching between two images which have large tile is also a difficult point. Aim at this problem,the SIFT and ASIFT algorithm are studied, besides, the characteristic of aerial images is considered,revise tile angle and translation variable of the Imaging Model in ASIFT algorithm. Partial solution to the problem of automatic matching of aerial images which have large tile, and bring out the experimental results.

Since 1995 when wave-front coding was first proposed by Dowski and Cathey, this technique has become a powerful tool to extend the DOF (depth of field) of incoherent imaging systems. By introducing a phase mask to the aperture plane, the optical transfer function (OTF) can be made insensitive to defocus and thus the key to this technique lies in the design of suitable phase masks. So far, many phase masks have been proposed to achieve the goal of DOF extension. In this paper, a new kind of phase mask, called composite phase mask, is designed by combining existing phase masks. With the optimized parameters, the composite phase mask can be made even less sensitive to defocus than any single mask. At the same time, as the number of phase masks used to generate one composite phase mask increases, the DOF extension effect can also be changed. Both the simulations and experiments demonstrate the effectiveness of the composite phase masks and the work reported also enriches the phase mask family used to realize the large DOF effect.

Affine invariant image comparison is always consequential in computer vision. In this paper, affine-SURF (ASURF) is introduced. Through a series of affine transformations and feature extraction, the matching algorithm becomes more robust with the view and scale change. A kd-tree structure is build to store the feature sets and BBF search algorithm is used in feature matching, then duplicates are removed by the conditional of Euclidean distance ratio. Experiments show it has a good result, comparisons with SIFT and SURF is made to prove its performance.

Real world data is usually high dimensional, and dimensionality reduction can significantly improve the efficiency of data processing and analysis. Existing approaches relying on distances between neighboring features typically suffer from the unreliable estimation of the true distance on a feature manifold due to its non-convexity. An approach is proposed to solve the problem by discarding long geodesics poisoned by boundary points indiscriminately. However, despite the improved performance, there are two major shortcomings with the approach. First, many long geodesics poisoned by few boundary points, which contribute little to the distortion of a manifold, are thrown away, as may decrease the robustness without improving the distortion of the manifold. Second, since short geodesics are sensitive to noise, retaining the whole effect of them may result in the bad robustness. This paper presents a regularization framework for nonlinear dimensionality reduction that incorporates long geodesics poisoned by few boundary points and reduces the effect of short geodesics, to realize isometry largely. In addition, the approach is sensitive to non-uniform sampling. To cope with the issue, we describe an improved robust boundary detection method. Experimental results are presented to illustrate the better performance of the proposed algorithm on two standard data sets.

In this paper, a novel radiometric compensation algorithm based on Lambertian reflection model is proposed to neutralize the visual perception of colors and textures which are intrinsic to the projection display surface. The algorithm uses a calibration method of advanced coded structure light to determine the geometric mapping between corresponding points in the projector-camera system. Concrete analysis of the dynamic range of projectors is applied to obtain the intensity limits of the input image. Consequently, the Lambertian reflection model is constructed for each point to compute the radiometric compensation function between displayed images and camera captured images. Experimental results show that this algorithm can effectively correct color inaccuracy of projection on arbitrary textured surfaces, and observers almost hard to notice visible artifacts of projected images.

Hand gesture recognition is receiving more and more attentions due to its potential use in many applications. In this paper, we propose a novel gesture spotting and recognition method, which combines the information of hand motion parameter, the matching result of HMM models and the recognition result based on geometrical features of hand trajectory to spot and recognize the gesture. Besides, we also study the method of adjusting classifiers to make the gesture recognition system adapt to specific users. Experimental results have proved the effectiveness of the proposed method.

In this paper, a strategy of reconstructing high resolution facial image based on that of low resolution is put forward. Rather than only relying on low resolution input image, we construct a face representation dictionary based on training high resolution facial images to compensate for the information difference between low and high resolution images. This restoration is realized through enrolling a low resolution facial image dictionary which is acquired through directly downsampling the learned high resolution dictionary. After the representation coefficient vector of a low resolution input image on low resolution dictionary is obtained through ℓ1-optimization algorithm, this coefficient can be transplanted into high resolution dictionary directly to restore the high resolution image corresponding to input face. This approach was validated on the Extended Yale database.

Obtaining high-resolution images from low-resolution ones has been an important topic in computer vision field. This is a very hard problem since low-resolution images will always lose some information when down sampled from high-resolution ones. In this article, we proposed a novel image super-resolution method based on the sparse assumption. Compared to many existing example-based image super-resolution methods, our method is based on single original low-resolution image, i.e. our method does not need any training examples. Compared to other interpolation based approach, like nearest neighbor, bilinear or bicubic, our method takes advantage of the inner properties of high-resolution images, thus obtains a better result. The main approach for our method is based on the recently developed theory called sparse representation and compress sensing. Many experiments show our method can lead to competitive or even superior results in quality to images produced by other super-resolution methods, while our method need much fewer additional information.

We use Fourier descriptors (FD's) to represent shape in model-based pose estimation. Specific invariance and normalization requirements for shape descriptors concerning to pose estimation are proposed. FD's are improved to meet such requirements. Common issues and techniques for such application are focused on. Starting point of the shape contour is fixed to the upper left corner point. Distance pairs instead of edge coordinate pairs recast the shape sequence. Moving average filtering is proceeded to remove the noises of the shape sequence. Shape sequence is re-sampled to make it definite length. FD's amplitude is normalized to the range of 0 to 1. Variance of sequences between observed and library FD's is defined as the shape matching objective function. For simulation, we use a missile model of Milkshape3d format. Results show that by adopting the improved FD's, we can arrive at a pose estimation practically by randomly optimal search of library FD's.

Skin color has been used as an important cue for various human related computer vision applications. However, detecting skin colors under varying illumination is a challenging task, as the appearance of skin in an image highly depends on the illumination under which the image was taken. To this end, a method for detecting skin colors under varying illumination is proposed in this paper. First, spatial illumination variation is identified and the images are segmented into different regions with different illumination. Each illumination region of color images are corrected base on the illuminant estimated by a local edge-based color constancy algorithm. Then, the corrected images are transformed into a color-space, where statistical results on a skin dataset show that the skin color cluster and non-skin color clusters are separated. Finally, the skin colors are modeled under Bayesian decision framework and classified from non-skin colors. The experimental results show that the proposed method is robust to illumination variations.

Frost is a kind of ground coagulation phenomena, and if the temperature of dew point is below 0C^o , the water vapor condenses as solid, which is called frost. The frost phenomena observing is an important step in daily ground observation work, and the results is one of 36 critical data in meteorological observation field. This work is usually accomplished by manual. In this paper, we propose an effective method for frost observation based on image processing. The changing of frost formation process is well simulated by using the curve fitting of gray correlation coefficient between certain lengths of frames, while the characteristic of frost surface texture is also well described by texture analysis based on texture descriptor. The experiment results show that our method can get high detection accuracy in the different kinds of continuous changing environment.

In this paper, a novel automatic DSM and remote sensing images registration scheme using template matching technique is developed. Due to the heterogeneity of DSM and remote sensing images, the emphases of our scheme are to describe the common feature between DSM and remote sensing images, and to generate a suitable template for template matching. Based on the sparse representation theory, we present a new feature descriptor, which can highlight the similarities of DSM and remote sensing images, and can be used to form a new kind of feature image. Meanwhile we present a criterion to choose the proper region from the feature image as the template which will ensure perfect template matching performance. The experiment results show that our scheme is efficient to fulfill the task of registration.

The concept of invariant moment of gradient orientation histogram and a novel line matching algorithm based on the invariant moment of histogram are proposed to resolve the problem of matching typical objects in the IR and visible images. First, line segments are extracted. Second, the average gradient vector direction of each pixel on the line is adopt as the main direction of the line. Third, the line is divided into non-overlapped sub-regions with the same size. The gradient vector direction of each sub-region are constructed, and the weighted invariant moment of histogram in each sub-region are calculated to build the line descriptor. Finally, the feature matching is realized via the NNDR(nearest/next ratio) method. Experimental results show that the proposed algorithm can match the typical objects in the IR and visible images efficiently and correctly.

An advanced user-oriented summary extraction method for soccer video is proposed in this work. Firstly, an algorithm of user-oriented summary extraction for soccer video is introduced. A novel approach that integrates multimodal analysis, such as extraction and analysis of the stadium features, moving object features, audio features and text features is introduced. By these features the semantic of the soccer video and the highlight mode are obtained. Then we can find the highlight position and put them together by highlight degrees to obtain the video summary. The experimental results for sports video of world cup soccer games indicate that multimodal analysis is effective for soccer video browsing and retrieval.

This paper developed a software platform of intersecting line cutting machine. The software platform consists of three parts. The first is the interface of parameter input and modify, the second is the three-dimensional display of main tube and branch tube, and the last is the cutting simulation and G code output. We can obtain intersection data by intersection algorithm, and we also make three-dimensional model and dynamic simulation on the data of intersecting line cutting. By changing the parameters and the assembly sequence of main tube and branch tube, you can see the modified two-dimensional and three-dimensional graphics and corresponding G-code output file. This method has been applied to practical cutting machine of intersecting line software.

A laser beam based calibration method is proposed in the paper. The laser beam we used can keep the cylinder shape in the real world space. The cylinder shaped laser beam will form an elliptical area in the plane in which the defect we want to measure also located. This elliptical laser area will form an ellipse in the image. Our purpose is to achieve calibration only depending on the elliptical parameters of the ellipse in the image. We created the relationship model between the elliptical parameters of the ellipse in the image and the calibration parameters of the video appliance, and we find a quick and easy way to solve the relationship model. Using this calibration method we can achieve contactless video measurement during the equipment checking in the nuclear power station.

A new method for circle detection, Hough gradient clustering method, has been developed in this paper. By using gradient direction angle to find the diameter of a circle, the new method can rapidly detect the circle in a complex background. The crucial steps in this method are the feature extraction and the clustering of the points which have the same gray direction angle and are collinear along the gradient direction. The application of the two-to-one space mapping and 1-2Hough transform can further reduce the useless calculation in the process of circle detection. Comparing with the Hough gradient method in OpenCV, the newly developed method shows a higher efficiency of circle detection in a complex background image as well as a great improvement in the anti-noise ability.

This article presents a method for the object classification that combines a generative template and a discriminative classifier. The method is a variant of the support vector machine (SVM), which uses Multiple Kernel Learning (MKL). The features are extracted from a generative template so called Active Basis template. Before using them for object classification, we construct a visual vocabulary by clustering a set of training features according to their orientations. To keep the spatial information, a "spatial pyramid" is used. The strength of this approach is that it combines the rich information encoded in the generative template, the Active Basis, with the discriminative power of the SVM algorithm. We show promising results of experiments for images from the LHI dataset.

Real-time stereo vision systems have many applications such as automotive and robotics. According to the Middlebury Stereo Database, Semi-Global Matching (SGM) is commonly regarded as the most efficient algorithm among the top-performing stereo algorithms. Recently, most effective real-time implementations of this algorithm are based on reconfigurable hardware (FPGA). However, with the development of General-Purpose computation on Graphics Processing Unit, an effective real-time implementation on general purpose PCs can be expected. In this paper, a real-time SGM realization on Graphics Processing Unit (GPU) is introduced. CUDA, a general purpose parallel computing architecture introduced by NVIDIA in November 2006, has been used to realize the algorithm. Some important optimizations according to CUDA and Fermi (the latest architecture of NVIDA GPUs) are also introduced in this paper.

In this paper, we First makes selection of the Zernike moment features of handwritten numerals based on the principles that the distinction degree of inside-class features is small and the dividing of the features between classes is huge; Then construct the similar matrix between handwritten numerals by the similarity measure based on Grey relational analysis and make transitivity transformation to similar matrix for better block symmetry after reformation; Finally make spectrum decomposition to the Laplacian matrix which from the reformation similar matrices, and recognize the handwritten numerals with the eigenvectors corresponding to the second minimal eigenvalues in Laplacian matrix as the spectral features. The experimental result indicates that the robustness of the algorithm proposed in this paper is great and the result is fine.

Monitoring urbanization may help government agencies and urban region planners in updating land maps and forming long-term plans accordingly. In this paper, a novel method for fast extracting residential area from remote sensing images based on log-Gabor filter was proposed. The method is divided in three steps. Firstly, we detect the edge-oriented urban characteristics in a remote sensing image using log-Gabor filter. Secondly, with the filtering orientations perpendicular to each other, we choose two log-Gabor filter response images to suppress the noise and acquire a smooth spatial region. Thirdly, a set of smooth regions served as residential areas can be extracted using Otsu's method. We tested it on diverse aerial and satellite images and encouraging results were acquired. The comparison of our method with the classical texture analyzing method of co-occurrence matrix demonstrated its superiority.

Tracking deforming objects involves estimating the global motion of the object and its local deformations as a function of time. Tracking algorithms using Graph Cut have been proposed. However, because of its globally optimal nature, it is prone to capture outlying areas similar to the object of interest. Constraint was introduced to confine the standard graph cut technique in many methods. In this paper, narrow band is introduced to constrain the standard graph cut where it is build on. With the narrow band constraint, segmentation is ensured to remain in it, which is useful when the result is not robust using standard graph cut. Secondly, Graph Cut is built based on the band region, compared with traditionally method, calculation cost is reduced. Thirdly, formulations are adjusted to distinguish between different parts of the narrow band, which improve the segmentation quality.

Tangent distance method approximates nonlinear manifolds by their tangent hyperplanes and has been widely used in image recognition. However tangent distance directly deals with original images while high-order statistic information may be neglected. And the information of image transformation should be known a priori. We propose a new image distance metric-Gabor feature-based approximated manifold distance (GFMD) to address these disadvantages. Firstly Gabor wavelet transform are applied to calculate high-order statistical information of images. The intrinsic variables of feature manifold are revealed by MVU. The feature manifold can be approximated by curve surfaces based on second-order Taylor expansion. GFMD is defined as the minimum distance between the approximated curved surfaces and can be directly combined with distance-based classifiers for image recognition. The experimental results of face recognition demonstrate that GFMD not only has higher invariance of transformation but also has more stability of classification than several state-of-the-art distance metrics.

Based on the variational idea, we propose a new fusion strategy for nonsubsampled contourlet transform (NSCT). For NSCT bandpass subband coefficients of input images, we take the main component of coefficients as the target and then build an extremum problem for energy functional to find the closest to the target one as the fused coefficient. We apply the gradient descent flow to minimize the functional and give the numerical scheme. The experimental results show that the proposed strategy outperforms state-of-the-art image fusion strategies for NSCT in terms of both visual quality and objective evaluation criteria.

3D reconstruction is one of main techniques for computer vision. A novel 3D convex surface reconstruction method is presented in this paper, which is based on visual hull principle. The real object is supposed to be surrounded by a 3D grid bounding box filled with voxels. A series of images of the object are captured by a calibrated camera in different locations. For each image, a series of virtual rays, each of which starts from an image silhouette point and drills through the camera center, intersect the voxels in bounding box to obtain a number of potential object surface points. Afterward, the potential surface points are projected onto the other images to eliminate the pseudo surface points which must locate outside the object image area in at least one image. When all images are processed, the surface points of whole object are obtained and then 3D surface is reconstructed. The experiment illuminates the feasibility and validity of our approach.

Biologists often use gene chip to get massive experimental data in the field of bioscience and chemical sciences. Facing a large amount of experimental data, researchers often need to find out a few interesting data or simple regulations. This paper presents a set of methods to visualize and analyze the data for gene expression signatures of people who smoke. We use the latest research data from National Center for Biotechnology Information. Totally, there are more than 400 thousand expressions data. Using these data, we can use parallel coordinates method to visualize the different gene expressions between smokers and nonsmokers and we can distinguish non-smokers, former smokers and current smokers by using the different colors. It can be easy to find out which gene is more important during the lung cancer angiogenesis in the smoking people. In another way, we can use a hierarchical model to visualize the inner relation of different genes. The location of the nodes shows different expression moment and the distance to the root shows the sequence of the expression. We can use the ring layout to represent all the nodes, and connect the different nodes which are related with color lines. Combined with the parallel coordinates method, the visualization result show the important genes and some inner relation obviously, which is useful for examination and prevention of lung cancer.

A novel labeling algorithm is proposed in this paper for high-resolution image connected components labeling. The proposed method successfully solves the labeling problem in the case that the computer memory capacity is less than required for the existing algorithms. Unlike the conventional algorithms that have to load the entire image into memory for labeling, our algorithm can finish the labeling task with the memory requirement of only two image rows size. Comparison and tests for high-resolution images are performed with the state-of-the-art algorithm. The proposed algorithm achieves significant performance improvement for high-resolution images labeling.

Automatically target contour detection from cluttered scenes is a very difficult task for computer vision. Humans, however, have a much better background suppress ability. The preceding models could not implement such a task very well. In this letter, an effective contour integration method based on human visual perception mechanism is proposed. The algorithm combines the properties of primary visual cortex and psychology researching results to simulate the contour perception of the V1 cortex. The new lateral connection based computational model have a better texture suppress ability, while, target's contour is enhanced. Compared with traditional methods, experiments show that the new method implement a more reasonable simulation of the V1 function structure, availably enhance the target's contour while suppress the cluttered background, obtain a balance between over and lose detection, besides, it has better accuracy with less computational complexity and time-consuming.

With the swift and violent development of the Multimedia Messaging Service (MMS), it becomes an urgent task to filter the Multimedia Message (MM) spam effectively in real-time. For the fact that most MMs contain images or videos, a method based on retrieving images is given in this paper for filtering MM spam. The detection method used in this paper is a combination of skin-color detection, texture detection, and face detection, and the classifier for this imbalanced problem is a very fast multi-classification combining Support vector machine (SVM) with unilateral binary decision tree. The experiments on 3 test sets show that the proposed method is effective, with the interception rate up to 60% and the average detection time for each image less than 1 second.

The objective of this work is to retrieve shots containing particular people in video material. Due to well known problems such as illumination, pose and expression variations, it has been a challenging task. To this end, a robust and scalable approach analogized with text search method is proposed in this paper, which mainly consists of four steps: Obtaining sets of face exemplars by tracking, Local Facial Features Extraction and Fusion, Visual words established, Document descriptor obtained and VSM established for shot retrieval. Experimental results demonstrate that the proposed approach is not only efficient but also outperforms the competing methods.

Cloud recognition is the base of weather forecast and the recognition of cloud types is challenging because the texture of the clouds is extremely variable under different atmospheric conditions. In this paper, we propose a novel method for ground-based cloud classification. Firstly, the interest operator feature (IO) and the sorted spectral histogram (SSH) feature are generated from Gabor-filtered images and then they are selected by using the principal component analysis (PCA), which can reduce the feature's dimension. Secondly the new training set is selected using the supervised clustering technology. Finally we send the two features to the multi-class SVM classifier, and a voting algorithm is used to determine the category of each cloud. In practice, we find no single feature is best suited for recognizing all these classes. The result shows that this method has higher classfication accuracy and lower space complexity than the other methods.

An imaging system for the continuous casting mold copper surface is researched and developed, to replace the on-line manual measuring method, which is used to checking Copper defects such as wearing, scratches and coating loss and other phenomena. Method: The imaging system proposes a special optical loop formed by three Mirrors, selects light source, CCD camera and lens type, designs mechanical transmission system and installation platform. Result: the optical loop and light source can insure imaging large-format object in narrow space. The CCD camera and lens determine the accuracy of horizontal scanning, and the mechanical transmission system ensures accuracy of the vertical scan. The installation platform supplies base and platform for the system. Conclusions: CCD-based copper surface imaging system effectively prevent defects such as missed measuring and low efficiency, etc. It can automatically and accurately shoot copper surface images on-line, and supply basis for image processing, defects identification and copper changing in the late.

With the development of society, handwritten digits recognition technique has been widely applied to production and daily life. It is a very difficult task to solve these problems in the field of pattern recognition. In this paper, a new method is presented for handwritten digit recognition. The digit samples firstly are processed and features extraction. Based on these features, a novel immune network classification algorithm is designed and implemented to the handwritten digits recognition. The proposed algorithm is developed by Jerne's immune network model for feature selection and KNN method for classification. Its characteristic is the novel network with parallel commutating and learning. The performance of the proposed method is experimented to the handwritten number datasets MNIST and compared with some other recognition algorithms-KNN, ANN and SVM algorithm. The result shows that the novel classification algorithm based on immune network gives promising performance and stable behavior for handwritten digits recognition.

According to that the SIFT algorithm contains contour property in the process of computation of scale-space, this paper proposes an improved SIFT algorithm named Contour directed-SIFT to offset the flaws of real-time in SIFT algorithm. In order to improve the accuracy of SIFT algorithm, the paper improve the Contour directed-SIFT algorithm using the epipolar constraint. Taking the two aspects above, a new matching algorithm based on SIFT algorithm in binocular vision is proposed. This algorithm can assure the real-time and accuracy in matching images, which can be applied in matching of massive features of tridimensional vision.

Image artifacts caused by image deconvolution usually make people feel regretful. Especially for image deconvolution algorithm which is used for restoring motion blurred image. The artifacts removing is challenging for conventional image deconvolution approach. In this paper, we propose a novel method that removes artifacts from a deconvolved image which caused by image deconvolution. Firstly, blur method is used to get a blurred version of the original artifacts image. Secondly, gradient projection algorithm is used to adjust the gradient of the original image with the blurred image for reference. Finally, the Poisson solution is adopted to reconstruct the final high quality and artifacts free image. Moreover, the artifacts are also reduced at the Poisson interpolation stage. The experiments demonstrate that our algorithms are efficient.

This paper presents a simple and computationally efficient saliency extraction method for detecting dim small target from single frame in heterogeneous background. The proposed method is based on background subtraction (BS), which identifies targets from the portion of a image that differs significantly from a background model. A set of horizon-directional filters (HDF) with multi-scales are first implemented to effectively recover the background maps from the input image. As a result, the foreground maps are extracted by computing the absolute difference between the input image and the estimated background maps. Then the foreground maps are fused into the total saliency map using a simple scheme. Finally, the experimental results of various cluttered background images show that the proposed method is efficient and has an outstanding performance in dim small target detection just by thresholding the saliency map.

This paper proposed a design program of 3D crafts manufacturing system which is based on binocular stereo vision and plug-in technology. First, this paper gave an overall introduction to the system, including system structure, function and working flow. Second, this paper analyzed some key technologies in the system development. It proposed a 3D measure method of structured light based binocular stereo system, and introduced the structure of the measurement system and analyzed thoroughly the key technologies in 3D measure, including calibration of stereo system, stereo matching and so on. Besides, it did some researches on how to use Geomagic Studio to carry on point cloud processing and the STL reconstruction. Third, this paper analyzed how to develop the system software which mainly consists of 3D measure Geomagic plug-in and RE Geomagic plug-in using VC++ programming language and plug-in technology. At last, this paper carried on large numbers of experiments and simulation using the developed software and the platform built according to the development program proposed in this paper and received satisfactory results which have proven that the development program proposed in this paper is feasible and the developed software takes on usability and may be used in the development of RE and RP system.

The method of disturbing training data randomly to train individual classifiers has been widely applied in some ensemble learning methods such as Bagging and Boosting to achieve strong generalization ability, however, it seems something blind. In this paper, a new ensemble learning algorithm named GAVPEL is proposed. By using the hierarchy nature of the data set, GAVPEL leverages the generalized attribute value partitioning method to form an ensemble tree, called a generalized classifier hierarchy tree. While classifying, GAVPEL selects part of the individual classifiers based on attribute value and ensembles them with majority voting. Experiment results show that GAVPEL can efficiently improve generalization performance when compared with some popular ensemble learning algorithms.

In this paper, a novel feature extraction method for palmprint recognition termed as Two-dimensional Combined Discriminant Analysis (2DCDA) is proposed. By connecting the adjacent rows of a image sequentially, the obtained new covariance matrices contain the useful information among local geometry structures in the image, which is eliminated by 2DLDA. In this way, 2DCDA combines LDA and 2DLDA for a promising recognition accuracy, but the number of coefficients of its projection matrix is lower than that of other two-dimensional methods. Experimental results on the CASIA palmprint database demonstrate the effectiveness of the proposed method.

In this paper, a novel incremental learning method called online independent Lagrangian support vector machine (OILSVM) is proposed. It achieves comparable classification accuracy with benchmark Lagrangian support vector machine (LSVM), while still enjoying the time efficiency of online learning machines. As opposed to the newly proposed OLSVM that utilizes the KKT conditions as data selection strategy, the size of the solution obtained by OILSVM using a linear independence check is always bounded, which implies bounded memory requirements, training and testing time. Experimental results demonstrate the effectiveness of the proposed OILSVM.

Terrain reconstruction aims at acquiring height maps by detecting corresponding feature points from two or more down-looking remote sensing images. This paper proposes a terrain reconstruction algorithm based on epipolar line rectification and dense matching. At first, it uses fundamental matrix to rectify stereo images to make their epipolar lines parallel and remove the disparity in vertical direction. Then, dense matching based on grid method is applied, which can provide sufficient matching points to estimate disparity of the rectified images. Finally, the heights of the matched points can be calculated according to the obtained disparity and flight parameters. Experiments show that our algorithm can generate precise and reliable height maps for well depicting terrain features.

Straight line is an important intermediate symbol, and the straight line extraction is one of the key technologies in image processing and computer vision. Aiming to the characteristics of digital image spatial domain, a straight line model of image spatial domain is established, and a heuristic algorithm based on the model for straight line extraction is presented, Which gets the edge amplitude image by edge detecting, and extracts the straight line which matches the straight line model through employing heuristic search guided by the edge information. The simulation shows that the algorithm is robust for image noise, and it not only can extract the low contrast straight line, but also can eliminate the short fractured straight line effectively.

The visual detection based on structured light is put forward to solve the problem of image matching for stereo vision in recent years. It is paid much attention in the industrial environment, especially in the three-dimensional automatic detection and object recognition. In this paper, it is discussed about the accuracy of visual inspection techniques based on structured light. There are some factors of affecting measurement techniques based on structured light, which are object material, edge detection algorithms, uneven lighting, and so on. The experimental results reveal that these three factors have a great impact on the accuracy of the model measurement of irregular objects.

As high brightness of laser spot, it significantly distinguish from the other goals on objects surface. Therefore, in computer vision, laser spot often is applied as interaction tools, and an important measurement tool etc. In building large-scale multi-projection display system, accurate measurement of geometrical parameters is very important. Based on subdivision mesh, this paper presents laser spot tracking algorithm with sub-pixel precision. The experimental data validates this algorithm sub-pixel precision.

This paper puts forward an offline handwritten numeral recognition method based on background structural descriptor (sixteen-value numerical background expression). Through encoding the background pixels in the image according to a certain rule, 16 different eigenvalues were generated, which reflected the background condition of every digit, then reflected the structural features of the digits. Through pattern language description of images by these features, automatic segmentation of overlapping digits and numeral recognition can be realized. This method is characterized by great deformation resistant ability, high recognition speed and easy realization. Finally, the experimental results and conclusions are presented. The experimental results of recognizing datasets from various practical application fields reflect that with this method, a good recognition effect can be achieved.