This PDF file contains the front matter associated with SPIE Proceedings Volume 8919, including the Title Page, Copyright information, Table of Contents, Invited Panel Discussion, and Conference Committee listing.

Change detection of VHR (Very High Resolution) images is very difficult due to the impacts caused by the seasonal changes, the imaging condition, and so on. To address the above difficulty, a novel unsupervised change detection algorithm is proposed based on deep learning, where the complex correspondence between the images is established by Auto-encoder Model. By taking advantages of the powerful ability of deep learning in compensating the impacts implicitly, the multi-temporal images can be compared fairly. Experiments demonstrate the effectiveness of the proposed approach.

This paper presents a new feature matching algorithm for nonrigid multimodal image registration. The proposed algorithm first constructs phase congruency representations (PCR) of images to be registered. Then scale invariant feature transform (SIFT) method is applied to capture significant feature points from PCR. Subsequently, the putative matching is obtained by the nearest neighbour matching in the SIFT descriptor space. The SIFT descriptor is then integrated into Coherent Point Drift (CPD) method so that the appropriate matching of two point sets is solved by combining appearance with distance properties between putative match candidates. Finally, the transformation estimated by matching the point sets is applied to registration of original images. The results show that the proposed algorithm increases the correct rate of matching and is well suited for multi-modal image registration.

This paper addresses the problem of large scale content-based video retrieval with relevance feedback. We analyze the common methods which leverage local feature detectors to extract feature descriptors from video collections and perform multi-level matching after indexing and retrieval of feature vectors. Instead of learning similarity-preserving codes, an approach of relevance feedback in a light-weight way is proposed. A relevance model is proposed to merge semantic similarity with the original distance matching at descriptor level. By learning several weights using canonical correlation analysis (CCA), the resulting candidate list of similar videos changes according to relevance feedback. Finally, we demonstrate the improvement of the proposed method by experiments on a standard real world dataset.

Atherosclerotic lesions at the carotid artery are a major cause of emboli or atheromatous debris, resulting in approximately 88% of ischemic strokes in the USA in 2006. Stroke is becoming the most common cause of death worldwide, although patient management and prevention strategies have reduced stroke rate considerably over the past decades. Many research studies have been carried out on how to quantitatively evaluate local arterial effects for potential carotid disease treatments. As an inexpensive, convenient and fast means of detection, ultrasonic medical testing has been widespread in the world, so it is very practical to use ultrasound technology in the prevention and treatment of carotid atherosclerosis. This paper is dedicated to this field. Currently, many ultrasound image characteristics on carotid plaque have been proposed. After screening a large number of features (including 26 morphological and 85 texture features), we have got six shape characteristics and six texture characteristics in the combination. In order to test the validity and accuracy of these combined features, we have established a Back-Propagation (BP) neural network to classify atherosclerosis plaques between atorvastatin group and placebo group. The leave-one-case-out protocol was utilized on a database of 768 carotid ultrasound images of 12 patients (5 subjects of placebo group and 7 subjects of atorvastatin group) for the evaluation. The classification results showed that the combined features and classification have good recognition ability, with the overall accuracy 83.93%, sensitivity 82.14%, specificity 85.20%, positive predictive value 79.86%, negative predictive value 86.98%, Matthew’s correlation coefficient 67.08%, and Youden’s index 67.34%. And the receiver operating characteristic (ROC) curve in our test also performed well.

Visual saliency has recently attracted lots of research interest in the computer vision community. In this paper, we propose a novel computational model for bottom-up saliency detection based on manifold learning. A typical graphbased manifold learning algorithm, namely the diffusion map, is adopted for establishing our saliency model. In the proposed method, firstly, a graph is constructed using low-level image features. Then, the diffusion map algorithm is performed to learn the diffusion distances, which are utilized to derive the saliency measure. Compared to existing saliency models, our method has the advantage of being able to capture the intrinsic nonlinear structures in the original feature space. Moreover, due to the inherent characteristics of the diffusion map algorithm, our method can deal with the multi-scale issue effectively, which is crucial to any saliency model. Experimental results on publicly available data demonstrate that our method outperforms the state-of-the-art saliency models, both qualitatively and quantitatively.

Cloud is floating in the earth sky widely, irregularly and frequently. So it appears in the satellite imagery. The cloud in the remote sensing imagery especially high resolution remote sensing imagery and aerial image will largely reduce the remote sensing image quality and use ratio, hinder the further application and the subsequent processing. Cloud detection accurately is a necessary and important step in the remote sensing image data analysis processing. So, a new cloud detection method based on HSI color space and stationary wavelet transformation (SWT) according to the spectral properties of cloud and the different with other objects is proposed in this paper. First, transform the RGB to HSI of image; then SWT is implemented to achieve the low frequency; the last result of cloud detection is obtained by the segmentation and edge extraction use SOBLE. The experiments show that the approach can detect the cloud accurately, availably and quickly.

An improved method of three-dimensional (3D) reconstruction from Inverse Synthetic Aperture Radar (ISAR) sequences based on factorization are proposed in this paper, which can improve the accuracy of reconstruction, and increase the number of reconstructed 3D features. A segmentation method of feature points based on clustering analysis is applied, which can remove some false points from reconstructed 3D features to enhance the precision of three-dimensional reconstruction. The result of simulation images and real images show the validity of the algorithm.

Clustering is an effective mean for of marine environment data analysis. This paper proposes a clustering algorithm based on the “Velocity-Direction” histogram. First of all, the “Velocity-Direction” histogram is constructed based on the characteristics of marine environment vector field data. Secondly, the exact surface of histogram is reconstructed by the Gaussian kernel function to eliminate the contaminated data points in “Velocity-Direction” histogram. Finally, the FCM algorithm is introduced and modified for the “Velocity-Direction” histogram clustering. The initial number and clustering centers for the FCM algorithm are set as the local extremum in the constructed histogram surfaces. The experiment results based on the simulation and the NOAA marine environment vector field data verifies the effectiveness of the proposed algorithm.

Many existing bottom-up saliency detection methods measure the multi-scale local prominence by building the Gaussian scale space. As a kind of linear scale space, it is a natural representation of human perception. However the Gaussian filtering does not respect the boundaries of proto-objects and smooth both noises and details. In this paper, we compute the pixel level center-surround difference in a nonlinear scale space which makes blurring locally adaptive to the image regions. The nonlinear scale space is built by a efficient evolution techniques and extended to represent color images. In contrast to some widely used region-based measures, we represent feature statistics by multivariate normal distributions and compare them with the Wasserstein distance on l2 norm (W₂ distance). From the perspective of visual organization in imaging, many priors are proved to be efficient in global consideration. In order to further precisely locate the proper salient object, we also use the background prior as a global cue to refine the obtained local saliency map. The experimental results show that our approach outperforms 5 recent state of the art saliency detection methods in terms of precision and recall on a newly published benchmark.

The researches on calibration of star sensor rarely involve the exterior parameters and image distortion of the optical system. In order to get more accurate interior-exterior parameters of the optical system, this paper proposes exact calibration model and algorithm based on interior-exterior parameters. Basing on analyzing the imaging model of star sensor, the principle of the star sensor calibration is as follow: firstly, the two-step method is used to get the initial interior-exterior parameters; then Levenberg-Marquardt optimization algorithm is utilized to get the global optimal solution. Experiments show that the angular distance of stars can be reduced from 57" to 5.2" after calibration. In addition, the calibration method can effectively eliminate the coupling of the interior-exterior parameters, achieve higher measurement accuracy, and significantly improve the recognition rate of the star map.

Matching the template image in the target image is the fundamental task in the field of computer vision. Aiming at the deficiency in the traditional image matching methods and inaccurate matching in scene image with rotation, illumination and view changing, a novel matching algorithm using local features are proposed in this paper. The local histograms of the edge pixels (LHoE) are extracted as the invariable feature to resist view and brightness changing. The merits of the LHoE is that the edge points have been little affected with view changing, and the LHoE can resist not only illumination variance but also the polution of noise. For the process of matching are excuded only on the edge points, the computation burden are highly reduced. Additionally, our approach is conceptually simple, easy to implement and do not need the training phase. The view changing can be considered as the combination of rotation, illumination and shear transformation. Experimental results on simulated and real data demonstrated that the proposed approach is superior to NCC(Normalized cross-correlation) and Histogram-based methods with view changing.

The objective of this work is to recognize faces under variations in illumination. Previous works have indicated that the variations in illumination can dramatically reduce the performance of face recognition. To this end，an efficient method for face recognition which is robust under variable illumination is proposed in this paper. First of all, a discrete cosine transform(DCT) in the logarithm domain is employed to preprocess the images, removing the illumination variations by discarding an appropriate number of low-frequency DCT coefficients. Then, a face image is partitioned into several patches, and we classify the patches using Sparse Representation-based Classification, respectively. At last, the identity of a test image can be determined by the classification results of its patches. Experimental results on the Yale B database and the CMU PIE database show that excellent recognition rates can be achieved by the proposed method.

Statistics of the number of students in the classroom is very important for class surveillance. It can help teacher count the number of students and help students choose class for self-study. While as a canonical pattern recognition problem, it’s very difficult due to various appearances of students and other outliers such as bags and books. We want to find a good solution to this problem. A novel method for texture feature extraction is now proposed based on that difference of Frequency spectrum image belongs to different seat image. Regarding frequency spectrum image as the texture image, the texture characteristics which can represent those differences are extracted using texture analysis's method. At the same time, we combine the Local binary patterns feature with the texture characteristics to describe the texture of seats. Experiments on a real classroom dataset demonstrate that the accuracy of the proposed method reaches 91.3%.

Face recognition under changing lighting conditions and facial expression are a challenging problem in computer vision. The variations in illumination and facial expressions can dramatically reduce the performance of face recognition. In this paper, an efficient method for face recognition which is robust under illumination and facial expressions variations. The core of the algorithm based on dense correspondence which we used is characterized by LBP and regional gradient between images. Our experiment on the AR databases and ORL face databases, ORL databases as a supplement in this framework. The results show that the proposed approach is not only efficient but also outperforms the comparative methods.

Rice yield estimation is an important aspect in the agriculture research field. For the rice yield estimation, rice density is one of its useful factors. In this paper, we propose a new method to automatically detect the rice density from the rice transplanting stage to rice jointing stage. It devotes to detect rice planting density by image low-level features of the rice image sequences taken in the fields. Moreover, a rice jointing stage automatic detection method is proposed so as to terminate the rice density detection algorithm. The validities of the proposed rice density detection method and the rice jointing stage automatic detection method are proved in the experiment.

Most of recent work on action recognition in video employ action parts, attributes etc. as mid- and high-level features to represent an action. However, these action parts, attributes subject to some aspects of weak discrimination and being difficult to obtain. In this paper, we present an approach that uses mid-level discriminative Spatial-Temporal Volume to recognize human actions. The spatial-temporal volume is represented by a Feature Graph which is constructed beyond on a local collection of feature points (e.g., cuboids, STIP) located in the corresponding spatial-temporal volume. Firstly, we densely sampling spatial-temporal volumes from training videos and construct a feature graph for each volume. Then, all feature graphs are clustered using spectral cluster method. We regard feature graphs as video words and characterize videos with the bag-of-features framework which we call it the bag-of-feature-graphs framework. While, in the process of clustering, the distance between two feature graphs is computed using an efficient spectral method. Final recognition is accomplished using a linear-SVM classifier. We test our algorithm in a publicly available human action dataset, the experimental results show the effectiveness of our method.

The detection of shadow is the first step to reduce the imaging effect that is caused by the interactions of the light source with surfaces, and then shadow removal can recover the vein information from the dark region. In this paper, we have presented a new method to detect the shadow in a single nature image with the saliency map and to remove the shadow. Firstly, RGB image is transferred to 2D module in order to improve the blue component. Secondly, saliency map of blue component is extracted via graph-based manifold ranking. Then the edge of the shadow can be detected in order to recover the transitional region between the shadow and non-shadow region. Finally, shadow is compensated by enhancing the image in RGB space. Experimental results show the effectiveness of the proposed method.

Higher quality surface information would be got when data from optical images and LiDAR were integrated, owing to the fact that optical images and LiDAR point cloud have unique characteristics that make them preferable in many applications. While most previous works focus on registration of pinhole perspective cameras to 2D or 3D LiDAR data. In this paper, a method for the registration of vehicle based panoramic image and LiDAR point cloud is proposed. Using the translation among panoramic image, single CCD image, laser scanner and Position and Orientation System (POS) along with the GPS/IMU data, precise co-registration between the panoramic image and the LiDAR point cloud in the world system is achieved. Results are presented under a real world data set collected by a new developed Mobile Mapping System (MMS) integrated with a high resolution panoramic camera, two laser scanners and a POS.

In this paper, an approach for the similarity-based global optimization of buildings in urban scene is presented. In the past, most researches concentrated on single building reconstruction, making it difficult to reconstruct reliable models from noisy or incomplete point clouds. To obtain a better result, a new trend is to utilize the similarity among the buildings. Therefore, a new similarity detection and global optimization strategy is adopted to modify local-fitting geometric errors. Firstly, the hierarchical structure that consists of geometric, topological and semantic features is constructed to represent complex roof models. Secondly, similar roof models can be detected by combining primitive structure and connection similarities. At last, the global optimization strategy is applied to preserve the consistency and precision of similar roof structures. Moreover, non-local consolidation is adapted to detect small roof parts. The experiments reveal that the proposed method can obtain convincing roof models and promote the reconstruction quality of 3D buildings in urban scene.

In this paper, a hybrid approach is proposed to detect texts in natural scenes. It is performed by the following steps: Firstly, the edge map and the text saliency region are obtained. Secondly, the text candidate regions are detected by connected components (CC) based method and are identified by an off-line trained HOG classifier. And then, the remaining CCs are grouped into text lines with some heuristic strategies to make up for the false negatives. Finally, the text lines are broken into separate words. The performance of the proposed approach is evaluated on the location detection database of ICDAR 2003 robust reading competition. Experimental results demonstrate the validity of our approach and are competitive with other state-of-the-art algorithms.

Image rectification reduce the search space from 2-dimension to 1-dimension and improve the searching efficiency of stereo matching algorithm greatly. In this paper, a simple and convenient method, which fully considered image sequence of monocular motion vision, is proposed to rectify the calibrated image sequence. The method is based on coordinate system transformation, which can avoid the mass and complex computations, and the method rectifies image sequence (three images) at once, which is efficient in image sequence processing. In this method, the rectification is composed of several steps. Firstly, we establish a reference coordinate system by three movement position. The Z axis of the reference coordinate system o_XYZ is the normal vector of the plane which three positions located. The direction of X axis coincides with the baseline from position 2 to position 1. We set Y axis according to right-hand principle. Secondly, we set the x axis and z axis of reference image space coordinate system o_xyz coincides with the X axis and Z axis of the reference coordinate system, and the y axis is set to coincide with the line from position 2 to position 3. Finally, we deduce a homography matrix to realize the image rectification. Both image data and computer simulation data show that the method is an effective rectification method.

We present a novel method to construct low-dimensional linear discriminative subspaces in this paper. Our method is simple and the calculation cost is little. The new subspace we construct is low dimensional while retaining discriminative information of original feature space. This means that we can make full use of discriminative information both in original ranks space and original null space by constructing a low-dimensional subspace and its discriminative matrix. The performance achieved by our method shows its great potential in resolving image classification problems.

Ship detection based on video is important in the application of surveillance and marine safety, the detection results of tradition methods, such as background subtraction, have much noise because of background noise such as ocean wave. In this paper we present a simple but efficient method for ship detection, It is based on the edge information of single image and movement information of multi images. Firstly, detect those movement pixels used the background subtraction to the video image, and the distance transformation is operation on the difference images; Secondly, we detect the edge of video image used Canny detector , and morphological operation on the edge image, lastly, eliminate the movement pixels if their distance transformation value is bigger than the threshold. The experimental results demonstrate that is efficient to eliminate the background noise and detect the real target.

In order to fuse highly conflicting evidence effectively, a novel combination method based on weighted distance of evidence is proposed by taking the ideas of Murphy’s averaging method and Deng’s weighted averaging method. Firstly, the essentiality of each element in the frame of discernment is given by Murphy’s idea. Secondly, the weighted averaging distance between any two bodies of evidence(BOEs) is calculated under the modified City Block distance norm, further the support degree of each evidence supported by other evidences can be obtained. Thirdly, the normalized total support degree of each evidence is considered as the weights of BOEs, and a new weighted averaging BOE will be gained. Finally, the information fusion process can be realized by using the Dempster’s rule of combination. Simulation results show that the proposed method can deal with the highly conflicting evidence with better performance of convergence, and it also can recognize the target more effectively and fleetly.

This paper presents an algorithm for the automatic segmentation of indoor videos into foreground and background layers. Segmenting foreground from an indoor video with local foreground motion and illumination changes is challenging. We first detect key frames with reliable motion using nonparametric model in chromaticity space. From these key frames, we learn an appearance model as a color degenerating model. Robust indoor video segmentation is achieved by combining these learned color and structure cues in a Markov random field framework. Experimental results on different sequences demonstrate the effectiveness of our algorithm.

Automatic gender recognition based on face images plays an important role in computer vision and machine vision. In this paper, a novel and simple gender recognition method based on face geometric features is proposed. The method is divided in three steps. Firstly, Pre-processing step provides standard face images for feature extraction. Secondly, Active Shape Model (ASM) is used to extract geometric features in frontal face images. Thirdly, Adaboost classifier is chosen to separate the two classes (male and female). We tested it on 2570 pictures (1420 males and 1150 females) downloaded from the internet, and encouraging results were acquired. The comparison of the proposed geometric feature based method and the full facial image based method demonstrats its superiority.

With the rapid development of technologies, road traffic surveillance tends to be more intelligent. Detection of non-public vehicles driving in public bus lanes is one of the emerging applications. Commonly, fixed cameras are adopted in video surveillance systems. Compared with the limited monitoring areas of fixed cameras, mobile cameras can follow the moving targets and in this way greatly extend the monitoring areas. However, for mobile cameras, many detection methods do not perform well because the background is rapidly changing and the target is moving fast as well. In this paper, we propose a novel method to detect non-public vehicles driving in the bus lanes (hence violating the traffic regulations) using mobile cameras installed on buses. In particular, we first use Hough transform and SVM classifier with color features to detect bus lanes, and then use AdaBoost cascade classifier with Haar features to detect license plates in the bus lane area. Finally another SVM classifier is used to classify the color of the license plate to determine if it belongs to a non-public vehicle. As shown in the experiments, our method is proven to be robust to complex background and performs well in the real world situations.

Currently there is no algorithm which can be adapted to all of the imaging conditions. So, it is necessary for us to find a method to evaluate the existing ATR (automatic target recognition) algorithm. We do some researches on ATR algorithm performance evaluation based on test methodology. The basic idea of the algorithm performance evaluation is to establish the relationship model between the image quality characteristics and the algorithm’s performance. In this paper, the algorithm performance evaluation’s techniques are studied, which include the algorithm performance assessment framework, the universal test image database’s creating, and the research of the image quality evaluation model. Firstly, under the guidance of the orthogonal experimental design method, we construct a universal test image database which includes the simulation image and the outfield flight data. And then this paper propose a method to establish the relation model between image quality characteristic and ATR algorithm based on SVM classifier. Finally we use the model to evaluate algorithm’s performance. We conduct experiments on the matching algorithm’s performance evaluation. The experimental results show that the proposed evaluation framework is efficient and the evaluation model is well.

To tackle the problem that classic RANSAC (Random Sample Consensus) is limited by the assumption that a single model accounts for all of the data inliers, an algorithm of multi-planar-feature fitting from 3D point cloud based on BaySAC algorithm (Bayes Sample Consensus) is proposed (called multiBaySAC). First, as the mathematical models of most of primitives to be fitted are determinate, a statistical algorithm of hypothesis model parameters histogram is proposed to detect potential planar features. Instead of assuming constant prior probabilities of data points and choosing initial data sets by random as RANSAC, we then implement a conditional sampling method -- BaySAC for robust parameters estimation of potential planar features, by computing the prior probability of each data point and updating the inlier probabilities using simplified Bayes’ rule. For the purpose of multiple feature fitting, the sequential application of the above procedure is implemented following the removal of the detected set of inliers. The proposed approach is tested with point cloud data of buildings acquired by RIEGL VZ-400 laser scanner. The results show that the proposed Multi-BaySAC can achieve high computation efficiency and fitting accuracy of multiple planar feature fitting.

In this paper, we propose an accurate shadow detection method via online self-modeling without tuning any feature threshold and manual labeling work. A primary classification is obtained from the fusion of classification results of a weak classifier like a low-value chromatic threshold technique and the online learned shadow generative model. Then object skeleton property and shadow’s spatial structure characters are considered to remove the camouflages and output the final classification result, the detected shadow pixels are used as training samples in the learning phase without manually labeling work. Online shadow model is learned by using Gaussian functions to fit the histograms of differential Hue, Saturation, and Intensity between background pixels and corresponding shadow pixels. Experiments indicate that the proposed method achieve both high detective and discriminative rates and outperform the approaches which need tuning thresholds when applied scene changes in accuracy and robustness.

Texture classification is very important in remote sensing images, X-ray photos, cell image interpretation and processing, and is also the active research areas of computer vision, image processing, image analysis, image retrieval, and so on. As to spatial domain image, texture analysis can use statistical methods to calculate the texture feature vector. In this paper, we use the gray level co-occurrence matrix and Gabor filter feature vector to calculate the feature vector. For the feature vector classification under normal circumstances we can use Bayesian method, KNN method, BP neural network. In this paper, we use a statistical classification method which is based on SVM method to classify images. Image classification generally includes image preprocessing, image feature extraction, image feature selection and image classification in four steps. In this paper, we use a gray-scale image, by calculating the image gray level cooccurrence matrix and Gabor filtering method to get feature extraction, and then use SVM to training and classification. From the test results, it shows that the SVM method is the better way to solve the problem of texture features for image classification and it shows strong adaptability and robustness for image classification.

Density-based clustering methods are usually more adaptive than other classical methods in that they can identify clusters of various shapes and can handle noisy data. A novel density estimation method is proposed using both the knearest neighbor (KNN) graph and a hypothetical potential field of the data points to capture the local and global data distribution information respectively. An initial density score computed using KNN is used as the mass of the data point in computing the potential values. Then the computed potential is used as the new density estimation, from which the final clustering result is derived. All the parameters used in the proposed method are determined from the input data automatically. The new clustering method is evaluated by comparing with K-means++, DBSCAN, and CSPV. The experimental results show that the proposed method can determine the number of clusters automatically while producing competitive clustering results compared to the other three methods.

The robust and rapid matching of oblique UAV images of urban area remains a challenge until today. The method proposed in this paper, Nicer Affine Invariant Feature (NAIF), calculates the image view of an oblique image by making full use of the rough Exterior Orientation (EO) elements of the image, then recovers the oblique image to a rectified image by doing the inverse affine transform, and left over by the SIFT method. The significance test and the left-right validation have applied to the matching process to reduce the rate of mismatching. Experiments conducted on oblique UAV images of urban area demonstrate that NAIF takes about the same time as SIFT to match a pair of oblique images with a plenty of corresponding points and an extremely low mismatching rate. The new algorithm is a good choice for oblique UAV images considering the efficiency and effectiveness.

Camera calibration is essential to obtaining three-dimensional information from two-dimensional image, this paper combines the method of photogrammetry and computer vision, put forward a kind of camera self-calibration based on hierarchical reconstruction and bundle adjustment. The projective reconstruction is obtained by SVD of the measurement matrix, Kruppa equation are deduced for calculating the camera parameters, then upgrade projective reconstruction to Euclidean reconstruction. Executing overall optimization to solve the inner orientation elements of the camera and the lens distortion parameters by bundle adjustment .Characteristics of this method is simple, not requested to build the field of high-precision control, just around the target for three or more images, the inner orientation elements of the camera and distortion parameters are solving ,achieving the camera self-calibration.

During the process of three-dimensional vision inspection for products, the target objects under the complex background are usually immovable. So the desired three-dimensional reconstruction results can not be able to be obtained because of achieving the targets, which is difficult to be extracted from the images under the complicated and diverse background. Aiming at the problem, a method of three-dimensional reconstruction based on the graph theoretic segmentation and multiple views is proposed in this paper. Firstly, the target objects are segmented from obtained multi-view images by the method based on graph theoretic segmentation and the parameters of all cameras arranged in a linear way are gained by the method of Zhengyou Zhang calibration. Then, combined with Harris corner detection and Difference of Gaussian detection algorithm, the feature points of the images are detected. At last, after matching feature points by the triangle method, the surface of the object is reconstructed by the method of Poisson surface reconstruction. The reconstruction experimental results show that the proposed algorithm segments the target objects in the complex scene accurately and steadily. What’s more, the algorithm based on the graph theoretic segmentation solves the problem of object extraction in the complex scene, and the static object surface is reconstructed precisely. The proposed algorithm also provides the crucial technology for the three-dimensional vision inspection and other practical applications.

In this paper, a novel classification method path-based similarity with instance-level constrains for SemiBoost, PBS-SB in short is proposed, and we exploit it for synthetic aperture radar automatic target recognition (SAR-ATR). Different from traditional SemiBoost method that uses the Gaussian kernel similarity, PBS-SB utilizes the path-based similarity, which considers the global consistence of data clusters. Besides, the instance-level constraints are integrated into the similarity measurement to construct the semi-supervised similarity, which provides the local consistence information. The experiments on 5 different data sets and MSTAR (Moving and Stationary Target Acquisition and Recognition) database demonstrate that the proposed method has superior classification performance with respect to competitive methods.

This paper presents our research on exploring the combined 2D TIN and 3D Tetrahedra structure to quickly model large-range Urban LIDAR point clouds for 3D visualization purpose. To this end, Morphological grayscale reconstruction is first implemented to segment LIDAR point clouds into terrain and non-terrain regions. After that, segmented Lidar terrain points are modeled with Constrained Delaunay Triangulation under constrain of building boundary as well as non-terrain points are modeled with Power Crust algorithm to obtain reconstructed building surface. Next, two kinds of model are combined based on shared building boundary. Finally, 3D visualization of selected urban area with presented technique clearly demonstrates higherefficiency. Valuable conclusions are given as well.

This paper presents our research on implementing SGM to obtain reliable edge points from intensity indoor image pair. To this end, the most basic theory of SGM is first outlined and with concern to the inherent nature of edge points, facts that take much advantage of SGM features are summarized as well. Next, we make discussion on detail procedures of edge matching with simplified SGM algorithm, in which gradient magnitude images is used to compute mutual information match cost and the matching costs of detected edge points are aggregated. Finally, the edge matching result with presented technique clearly demonstrates higher reliability and efficiency. Valuable conclusions are given as well.

To solve the problem that traditional HOG approach for human detection can not achieve real-time detection due to its time-consuming detection, an efficient algorithm based on first segmentation then identify method for real-time human detection is proposed to achieve real-time human detection in clutter scene. Firstly, the ViBe algorithm is used to segment all possible human target regions quickly, and more accurate moving objects is obtained by using the YUV color space to eliminate the shadow; secondly, using the body geometry knowledge can help to found the valid human areas by screening the regions of interest; finally, linear support vector machine (SVM) classifier and HOG are applied to train for human body classifier, to achieve accurate positioning of human body’s locations. The results of our comparative experiments demonstrated that the approach proposed can obtain high accuracy, good real-time performance and strong robustness.

Steam reheating system is emerging as a multivariable system with steam-steam exchanger, the strong coupling and time delay characteristics. The traditional approach for the predictive control in power plant requires modeling based on accurate mathematical model, and some multivariate statistical algorithm cannot avoid falling into the over-fitting, therefore these approaches is not suitable for prediction of the reheating temperature in power plants. In this paper, we used the least squares support vector machine (LS-SVM) regression algorithm to predict the temperature of the steam reheating in the power plant combined with the data set of the steam reheating in a 120MW power plant. Comparing with the existing algorithms, the result shows that the LS-SVM is a robust and reliable tool for prediction in engineering application field.

Images are one of vital ways to get information for us. However, in the practical application, images are often subject to a variety of noise, so that solving the problem of image denoising becomes particularly important. The K-SVD algorithm can improve the denoising effect by sparse coding atoms instead of the traditional method of sparse coding dictionary. In order to further improve the effect of denoising, we propose to extended the K-SVD algorithm via group sparse representation. The key point of this method is dividing the sparse coefficients into groups, so that adjusts the correlation among the elements by controlling the size of the groups. This new approach can improve the local constraints between adjacent atoms, thereby it is very important to increase the correlation between the atoms. The experimental results show that our method has a better effect on image recovery, which is efficient to prevent the block effect and can get smoother images.

The paper presents a gradient-based algorithm for image registration. The algorithm is extended from the classical Lucas-Kanade algorithm, and it aims to solve the rotation-scale-translation (RST) model. To solve the problem, the 6- parameter affine model is used, and the algorithm is derived according to the idea of the Lucas-Kanade algorithm, then the RST model parameters are obtained from the estimated affine model values. Due to its Taylor approximation nature, iterative scheme is needed, and the inverse compositional scheme by Keren et al. is used. To further increase the speed and convergence range, coarse-to-fine strategy is also used. In the final, simulations are performed to verify and evaluate the algorithm, and the results demonstrate that it can obtain sub-pixel estimation with high accuracy.

Biometrics recognition aims to identify and predict new personal identities based on their existing knowledge. As the use of multiple biometric traits of the individual may enables more information to be used for recognition, it has been proved that multi-biometrics can produce higher accuracy than single biometrics. However, a common problem with traditional machine learning is that the training and test data should be in the same feature space, and have the same underlying distribution. If the distributions and features are different between training and future data, the model performance often drops. In this paper, we propose a transfer learning method for face recognition on bimodal biometrics. The training and test samples of bimodal biometric images are composed of the visible light face images and the infrared face images. Our algorithm transfers the knowledge across feature spaces, relaxing the assumption of same feature space as well as same underlying distribution by automatically learning a mapping between two different but somewhat similar face images. According to the experiments in the face images, the results show that the accuracy of face recognition has been greatly improved by the proposed method compared with the other previous methods. It demonstrates the effectiveness and robustness of our method.

In this paper, a new multi-focus images fusion method based on wavelet transform is presented. Firstly, the source images are decomposed by wavelet transform. Secondly, the wavelet coefficients of approximate is computed as the significant sum of the corresponding coefficients at the same position in high frequency bands, region match of local energy is used to fuse high frequency sub-bands. Ultimately, the fusion image is obtained through an inverse wavelet transform. The fusion performance of proposed method has been evaluated through informal visual inspection and objective fusion performance measurements, experimental results show that the proposed method can achieve a better fusion performance compared to conventional fusion approaches. Moreover, the relation between wavelet base and image fusion is studied, which is of great value for research and experiment in this field.

Multi-resolution is the good characteristics of wavelet transform. In wavelet transform domain, high frequency subband of image is sparse. Less high frequency coefficient can be sampled by compressed sensing technology. In this study, for an image, a sparse representation in the wavelet transform domain is found. Image is reconstructed by the orthogonal matching pursuit (OMP) , compressive sampling matching pursuit (CoSaMP), iteratively reweighted least square (IRLS) and Suspace Pursuit (SP), respectively. Different wavelet basis and sampling rate which affect the quality of the reconstruction are discussed. Experimental result shows that performance of IRLS is the best, OMP are easy implementation and fast speed, and Coif3 has a better performance than the other wavelet basis.

Range image can be obtained by 3D-Scanning and needs registration. Based on the classic ICP (Iterative Closest Point) algorithm, this paper presents an improved ICP method. The classic ICP uses the 3D point-to-point distance as the error measurement function. In our paper, the point-to-point distance will be replaced by a point-to-facet distance. By formula derivation, this measurement function can be transformed into facet-weighted point-to-point distance. We apply this method for range image registration and the result shows the validity of this algorithm, which has faster convergence rate and better anti-noise attribute than previously described weighted ICP methods.