Advances in Image and Graphics Technologies

Viewpoints selection is a key part in 3D object recognition since 3D object can be represented by a set of 2D projections. In this paper, we discuss a new method to select viewpoints of 3D object recognition. Based on manifold topological multi-resolution analysis method (MMA), manifold information which represents the intrinsic feature of 3D objects is used, thus the viewpoints selected are more distinctive. We compared with “7 viewpoints method” which provides us a simple and effective way to select viewpoints. Experiments demonstrate that the method based on MMA is effective and performances better than “7 viewpoints method” in 3D object recognition.

The Phase Correlation(PC) method demonstrates high robustness and accuracy for measuring the very subtle disparities from stereo image pairs, where the baseline (or the base-to-height ratio) is unconventionally narrowed. However, this method remains inherently computationally expensive. In this paper, an adaptive PC based stereo matching method is proposed, aiming to achieve higher speed and better stereo quality compared to the existing methods, while also preserving the quality of PC. Improvement was achieved both algorithmically and architecturally, via carefully dividing the computing tasks among multiprocessors of the GPUs under a novel global-pixel correlation framework. Experimental results on our hardware settings show that the method achieves as high as 64× and 24× speedup compared to single threaded and multi-threaded implementation running on a multi-core CPU system, respectively.

Biological water quality monitoring is an emerging technology. The change of water quality can be quickly tested by using the sensitiveness of aquatic organisms to water environmental change. However, how to extract semantics of fish behavior from the video data is the key technical point of achieving water quality testing. On the base of quantifying fish behavioral characteristics, the essay puts forward the semantic descriptive model of fish behavior. By grouping the parameter of the amount and average height of multi-target fish movement and extracting the semantics of each group, the semantic descriptive network of fish behavior characteristics and water quality finally was set up. The experimental data show that the semantic descriptive network can better characterize the water quality in high temperatures. This provides the theoretical base for the applications of biological water quality tests by the behaviors of fish groups.

Infrared face recognition, being light- independent, and not vulnerable to facial skin, expressions and posture, can avoid or limit the drawbacks of face recognition in visible light. Local binary pattern (LBP), as a classic local feature descriptor, is appreciated for infrared face feature representation. To extract compact and principle information from LBP features, infrared face recognition based on LBP adaptive dominant pattern is proposed in this paper. Firstly, LBP operator is applied to infrared face for texture information. Based on the statistical distribution, the variable dominant pattern is attained for different infrared faces. Finally, dissimilarity metrics between the adaptive dominant pattern features is defined for final recognition. The experimental results show the adaptive dominant patterns in infrared face image have a lower feature dimensionality, and the proposed infrared face recognition method outperforms the traditional methods based on LBP uniform and discriminant patterns.

Single training sample face recognition problem is a challenge in face recognition field, so the distinguished feature extracting is important step for improving the recognition correct rate under the condition only one sample of one person in training set. Gabor feature and 2DPCA reducing dimension algorithm are also effective feature extracting method and are applied on face recognition and pattern analysis fields. But the two methods can’t be combined because that 2DPCA need inputting data with 2D structure. A feature extraction method based on Gabor and 2DPCA is proposed in this paper. It transforms a series of Gabor sub-images to an image with the help of image splicing technique, and then 2DPCA can be employed. Experimental results on ORL face dataset show that the proposed method is effective with higher correct rate than those of other similar algorithms for single face recognition.

As Visual Background Extractor (Vibe) is sensitive to illumination change and object shadow and extracts movement area incompletely, this paper proposes a Vibe motion target detection algorithm based on Lab color space. For matching pixels with their background models, the algorithm improves CIE1976Lab color-difference formula by reducing the proportion of brightness difference. Experiments show that the improved color-difference formula is less sensitive to illumination change and object shadow. Then, the algorithm makes use of space consistency of pixels to correct pixel classification results, which improves the anti-noise ability and makes extracted movement area more complete. Experimental results demonstrate that the algorithm proposed has better detection effects both in indoor and outdoor environments.

Steganography with high payload is the needs of the real application in the context of large data. Nowadays, the published steganography have the significant achievements on the embedding quality, but the embedding capacity usually is not enough. In this paper, a new image data embedding with large payload based on reference matrix method is proposed. The reference matrix is generated by a base-9 numeral system, which guides cover pixel pairs’ modifications. Compared to the recent methods, the experimental results show the proposed method not only ensures the accepted image quality, but has the larger embedding capacity which is up to 3.169 bpp.

A novel image forensic method based on generalized coefficient-pair histogram in DCT domain was proposed. In the proposed method, firstly, the image is transformed by DCT, and then the differential DCT coefficient matrix of two directions, such as horizontal and vertical direction are computed, the following is to compute the coefficient-pair histogram for each differential DCT coefficient matrix within the given threshold. Finally, support vector machine (SVM) is used to classify the authentic and spliced image through training the feature vectors of authentic and tampered image. The experimental results show that the proposed approach has not only the lower computing complexity; it also outperforms all the state-of-the-art methods in detection rate with the same test database.

Non-local means (NLM) is a powerful denoising algorithm that can protect texture effectively. However, the computational complexity of this method is so high that it is difficult to be widely applied in real-time systems. In this paper, we propose a fast NLM denoising algorithm which can product comparable or better result with less computation time than the traditional NLM methods. Some experimental results are provided to demonstrate the superiority of the proposed method.

This paper presents a new method of license plate location under complex background. The texture direction map was obtained by gradient direction field and the calculation of the original image. License plate candidate area is determined using the method of interval judgment edge information through texture direction and binary image. Finally, the plates are accurately positioned using the improved regional notation. This experiments results demonstrate the great robustness and efficiency of our method.

3D Reconstruction based on binocular data is significant to machine vision, and has the following basic steps: camera-self calibration, stereo matching, depth extraction, and 3D representation. Due to the high precision requirement, the configuration of dual camera in the binocular reconstruction system is often too strict to implement. And stereo matching as the main task can hardly be done well in both time computing and precision. Our study proposes a new and high efficiency processing flow, in which we use a consumer camera. The kernel feature is proposed in calibration stage to rectify the epipolar. The most prominent breakthrough is that we segment the objects in the camera into background and foreground, for which system obtains the disparity by different method: local window matching and kernel feature-based matching. Extensive experiments demonstrate our proposed algorithm represents accurate 3D model.

Based on hardware architecture of CUDA (Compute Unified Device Architecture), this paper not only makes full use of multithreaded and parallelism in GPU (Graphic Processing Unit, the image Processing Unit), but also takes advantage of the memory to improve parallelization of SIFT algorithm. What’s more, two-dimensional thread structure is adopted when storing the image data and variable blockIdx which is built in the device is used for mapping width and height in pixels of the two-dimensional images. Thus, it can improve the efficiency of parallel by making full use of thread parallelism and two-dimensional features of thread grid. The experimental results show that the matching accuracy and speed of the algorithm have greatly improved compared to the traditional serial SIFT algorithm, and the maximum acceleration ratio can reach 21.43 in this experiment making image parallel matching possible under the smart vehicle environment.

In order to get a better solution on non-linear and non-Gaussian targets tracking problems, a novel multi-object tracking framework based on particle filter with data association is proposed. Firstly, a self-adaptive size tracking window algorithm is given and integrated into the tracking framework for the changes scale of moving objects, then a novel data association method based on JPDA is proposed, in which the tracking window, the observation data of the same object at adjacent frames and the different objects at the same frame are jointly associated to achieve accurate data association during tracking. The simulation results indicates that the proposed framework can be used effectively in multi-object tracking, it have the ability of dealing with the challenges such as object occlusion, separating, merging, and pot up.

Recently, the helicopter patrol is the main way in power system line patrol with the advantages of high efficiency and low cost. The aerial images are characterized by large in number, complex background, et cetera. It is necessary to segment insulator object from aerial images for better insulators’ fault diagnosis. The traditional single segmentation method causes user’s fatigue and results in bad segmentation quality. This paper proposes a cosegmentation method of aerial insulator images which utilizes the relationship between images that can improve the segmentation quality and reduce user’s workload. According to the thermodynamic anisotropic diffusion theory and the constructed graph network, we extract the corresponding largest relevant region by temperature maximization among the images as the common insulator objects. In order to achieve more accurate and fast segmentation, we remove the text, noises in aerial images and over-segment the preprocessed images into superpixels. Experiments show that the method can obtain good results which are instrumental to insulators’ fault diagnosis.

With the popularity of smart mobile devices and the development of streaming media technology, the mobile video surveillance system is more and more popular. The Android mobile phone has been widely used because of open platforms, wide variety, good user experience, etc. In this paper, a solution for live video based on Android system is proposed. First, H.264 stream is unpacked based on the RTP protocol, Second, the generated package is transmitted via DatagramSocket (Socket of UDP protocols). Finally, dynamic library(.so) is generated from FFmpeg library by using Cygwin in the Windows system, while dynamic library to decode in the Android system. This paper focuses on the transmission and the decoding process of H.264 stream. The experimental results show the validity and practicability of the proposed solution and have met the demand of the programme.

The artificial counting method for solar wafers cannot meet the production requirements because of the inefficiency, high breakage rate and inaccuracy. An accurate counting algorithm for solar wafers based on the texture feature was suggested. Firstly, Texture feature of acquired multilayer solar wafers was analyzed. Secondly, the edges of wafers region were located in the image by the grayscale distribution. Coordinates information of edges were used to correct the image and locate ROI(Region-of-Interest). The thresholding combining NiBlack with background correction was designed to obtain binary image. Noise was removed by particle Filter. The texture feature gap between pieces was improved by morphological method. Finally, counting was realized by self-adaption differential algorithms based on probability statistics. The result of experiments shows that the counting accuracy has achieved to 99.37%. The algorithm mentioned above is not only effective for the wafer images of low contrast, much noise and fuzzy boundaries, but also improves the efficiency and accuracy of the counting.

Feature point matching is a key step of image registration, object recognition and many other computer vision applications. By using the proposed structural distance between feature point sets as the matching similarity, we are able to match the spatial structures of feature points in different images. In the optimization process of the structural distance, both local and global relationship are considered, which greatly improves the robustness and accuracy. We also present a fast algorithm with higher efficiency, which approximately realizes this method by linear matrix multiplication operations. The proposed method achieve promising matching results in the experiments.

This paper confirms the utility of Ohta color space, GLOM and MRF model to enhance the accuracy of segmentation of color textured images. The statistical properties of color textured images in Ohta color space are explored by means of GLOM and the segmentation is done by contextual modeling of the data through MRF modeling. The Haralick feature Mean at IPD 1, as optimized with this approach, appears to be the best textural feature to improve interclass discrimination. The results obtained by our tests are compared with those of MRF modeling in RGB color space and our method found to be the better choice.

In the structured light 3D shape measurement system, the projector plays an essential part for 3D shape reconstruction. As the projector calibration accuracy affects the 3D shape measurement accuracy, a flexible projector calibration approach in the structured light system is proposed. Key to the proposed method is to establish the relationship between the projector coordinate and the world coordinate using the pre-calibrated camera. The process of the system calibration can be divided into three steps. First, a black-and-white (B/W) checkerboard is pasted on the white board plane for camera calibration. Second, the pre-calibrated camera is used to obtain the 3D world coordinates of the checkerboard corners. Here, instead of using complicated process with spatial color illuminations or phase-unwrapping method for the pixel mapping, a projected B/W checkerboard is used for the projector calibration. Unlike some other methods, the non-wide-angle camera can be used to capture the images of both pasted and projected checkerboards. Also the calibration process is not vulnerable to the environment. Finally, the results of the camera and projector calibration are used to reconstruct 3D shape. Experiments show that the reprojection errors for both camera and projector are within ±1 pixel, and the RMS error of 3D reconstruction is 0.25 mm.

Distributed video coding is a coding paradigm that shifts the computational intensive motion estimation from encoder to decoder. The lightweight encoder is far more attractive for wireless sensor network and wireless video communication application. But there is seldom successful DVC application in the real world, because of the decoder is too complex to realize, also the performance of DVC is far from the H.264 codec. In fact as the development of the hardware, the mobile device is far stronger than before. In this paper, we propose a coding paradigm that makes the complexity of the encoder and decoder is comparable, which is called balanced video coding. The motion estimation of distributed video coding was moved from the decoder to encoder. The results of motion estimation computation were sent back to decoder. The experimental results show that the complexity of the encoder and decoder from balanced video coding was comparable.

Currently, the main obstacle hindering the development of gait recognition systems is the influence of changes of clothing, carrying goods and viewpoint on the gait profile of pedestrians. In this paper, we propose a gait recognition method based on Gravity Center Tracks (GCT) that converts the original video of the gait into a binary image sequence, calculates the coordinates of the gravity center in each frame of image, sequentially connects the coordinates of the gravity centers of all frames in the image sequence to obtain the GCT. The GCT is preprocessed and spectrum analyzed, and the transformed DFT coefficients are input into a K cluster and BP neural network for recognition after normalization. The test shows that the method is capable of achieving excellent recognition under conditions with different clothing and carrying goods and still maintains a high recognition rate without retraining models when the pedestrians change walking direction and walking conditions.

Nowadays, coupling Command and Control (C2) Systems with combat simulation systems becomes very important and data virtualization has not gained enough attention. The C2/M&S Interoperability Technical Reference Model (TRM) is briefly discussed, which divides data into three categories, viz. persistent data, non-persistent data and exercise control data. The heterogeneity of these data categories is analyzed and the component-based comprehensive data visualization approach for coupling C2 and combat simulation systems is proposed. The XTrain Federation is built to support commanders and staffs training which consists of the Army Tactical Command and Control System (ATCCS) and the Army Tactical Simulation System (ATSS). In this federation, ATSS generates dynamic track data that are exported to and displayed on the ATCCS as a common operational picture (COP); accordingly, train audiences inject actions/orders back into ATSS. Research results show data visualization plays a great role in the domain of coupling C2 with combat simulation systems.

Aiming at recognizing ship target efficiently and accurately, a novel method based on active learning and the Composite Kernel Support Vector Machines (CK-SVM) is proposed. First, we build a ship recognition dataset which contains the major warship models and massive civil ships. Second, to reduce the cost of manual labeling, active learning algorithm is run to select the most informative and representative samples to label. Finally, we construct a composite-kernel SVM combining shape and texture features to recognize ships. The composite-kernel strategy can enhance the quality of features fusion apparently. Experiments demonstrate that our method not only improves the efficiency of samples selection, but also receives satisfying results.

According to the iris theory, the iris is the organ which can show the function of various organs, so diseases can be diagnosed by the analysis of iris image. In this paper, the iris image is preprocessed, interference factors e.g. noise is removed, the edges inside and outside the iris image are located, the region which is concerned is extracted, plaques are extracted from the image and center coordinate is acquired according to the extraction method of gray threshold scope distance features based on the characteristics of the plaques, which will offer convincing theoretical support for automatically identifying the position of the disease by computer.

Compressed Sensing (CS) theory is a new method of signal acquisition and processing proposed in recent years. With small amount of sampling data recovering original data to precisely reconstruct sparse signal or compression signal, the theory breaks though the restriction of Nyquist sampling theorem. CS can avoid enormous sampling data waste but also reduce the complexity of image coding. This paper reviews the basic theory of CS and its three key points, including signal sparse representation, design of measurement matrix and reconstruction algorithms. Then, the application of CS in the field of remote sensing image compression technology is studied. Using MATLAB software, we do a series of CS emulation experiments compared with the traditional compression methods. The results show that the proposed method has a good performance on the remote sensing image compression.

Underwater drowning detection in public swimming pools is a challenging task. To detect drowning swimmers, an implementable real-time detection system with high accuracy is needed. In this paper, we propose a novel camera-based drowning detection algorithm. The input video sequences are obtained from under water cameras. Moving object in the alert zone will be extract from background by background subtraction. The inter-frame based denoising scheme is employed to eliminate complex interferences in the water. Experimental results are shown that the proposed algorithm can detect the drowning swimmer more accurately without massive computations.

Through analyzing the data collected by the 13-point based laser bar, the original rutting transverse profile is built in this paper. The angle rotation method and cubic spline interpolation method are then adopted to rotate and smooth the profile to get the standard figure of rutting transverse profile. The rotation processing eliminates the effects of crown slope on rutting profiles, and the smooth processing makes rutting profiles be more similar to the real pavement profile. Based on the comparison of straight-edge and wire-line rut depth calculation methods, the wire-line method is used to acquire the 1D rut depth and rut width. The trapezoidal integral rule is utilized to calculate the 2D rut positive area and negative area. Results show that the indicators extracted in this paper can describe the rut more comprehensively and accurately.

Minimally Invasive Surgery (MIS) has offered great benefits to the patients for reducing patient trauma and recovery time. However, traditional MIS still has some challenges due to the loss of 3D vision and the narrow field of view provided by the endoscope. In this paper, we propose a method to reconstruct real 3D model of gastric internal surface based on gastroscopic image sequences, which can supply a significant navigation for surgeons during MIS procedure. This method utilizes a novel six-degree of freedom (6-DOF) tracking sensor to record endoscope’s real-time position information, and Structure from Motion (SFM) theory is adopted to generate 3D gastric internal surface model. The system is experimented on simulated gastric model. The 3D depth accuracy is estimated at 9.8 mm, which is suitable for clinic practice. Theoretically, this system can be applied to MIS not only on stomach but also on other organs.

Liver segmentation from Computed Tomography (CT) image is important for the diagnosis and intervention of liver diseases. In this paper, we propose an automatic liver segmentation method based on probability image and random walks. First, pixel-level texture features are extracted and liver probability images are generated corresponding to the test images using a binary classification approach. Second, random walk algorithm with automatic seed points is developed to detect the liver region. The proposed method is validated on standard data with five evaluation criteria. Experimental results demonstrate the effectiveness and robustness of the proposed method for the liver segmentation in CT image. The proposed method can achieve an average volumetric overlap error of 8.76% and an average surface distance of 1.30 mm.

Traditional algorithm of Harris needs to select a parameter for computing interest values of pixels, and its recognition ability for some types of corners is poor. To solve this problem, this paper proposes a corner detection method which is based on local standard deviation and logarithmic computing. The method decreases effecting to the response values of corners that near the candidate interested points through computing the logarithms of gradient, so it can detect different types of corner more effectively. It can redefine the interest value function according to the statistical features of the standard deviation to decide the corners. The function could avoid selecting value of parameters by person, and it could directly judge whether a candidate interested point is a corner, to make the algorithm has a higher objectivity. The experimental results show that the method can effectively detect the corners of various types, and it can achieve a more accurate effect of positioning.

A single character cannot describe palmprint features accurately and affect recognition results. To solve this problem, a palmprint recognition method based on fusion of contour feature and edge texture information is proposed. Firstly, mean filter is used to decompose palmprint images to obtain low frequency layer and high frequency layer. Next, the block-based idea is utilized in the two layer images, where gray histogram is used to extract counter features from the low frequency layer, at the same time differential box counting is used to extract texture information from the high frequency layer. Then, these obtained features are fused in order to further improve recognition accuracy. Finally, a common Chi-square measure is used to measure the simlarity. Experimental results on PolyU palmprint database are compared with traditional palmprint recognition algorithm, the proposed method can obtain 99.56% recognition accuracy, the time of feature extraction and matching is only 55.641ms, the effectiveness of the method is proved by lower algorithm complexity and faster speed.

Splash-color painting is a unique painting style developed on the basis of Chinese traditional painting in recent years, it is the outstanding representative of the “freehand brushwork” in Chinese traditional painting. The splash-color creation with digital media will endow new vigor to the traditional painting art. The digital splash-color painting interactive system based on the mobile platform conducts the real-time extraction and parameterization for the handwriting and dressing colors of the painters, which not only simulates the natural creation process of traditional splash-color skills on the mobile platforms, but also expresses the subjective consciousness and subjective emotions highly-reflected in the process of creating splash-color paintings,which may further promote the spreading and development of traditional Chinese art in the digital age. Moreover, it is of positive explorative significance and practical value in the digital media design and planning of Chinese traditional painting art.

According to the detailed study of histogram and local optimal solution method which is similar with pyramid method, we proposed a novel video segmentation. The main contributions of our work are: 1) Calculating the histogram of video frames and getting the similarities of the adjacent frames by using three different methods which are correlation method, chi-square method and intersection method. 2) Drawing line graphs of these three similarity sequences. 3) Studying the characteristics of these three graphs, achieving local optimal solution which is based on the idea of drop sampling in pyramid to judge if the video scene has converted and segmenting the video. 4) The experimental results of the segment method are shown compared the three different distance methods at the step 1.

Tattooing and Argon plasma coagulation (APC) are used for gastroscopic biopsy traditionally. To overcome the invasive issues of tattooing and APC, we proposed an image-guided gastroscopic biopsy system (IGGBS) to guide endoscopist in retargeting previous biopsy sites in the follow-ups non-invasively. In this paper, a model merging method is proposed to improve the IGGBS’s accuracy. The method reconstructs local realistic model based on gastroscopic image sequences during procedure, then the local model is merged with the pre-operative model in real-time. As a consequence, the merging model is used in IGGBS to navigate endoscopist to retarget biopsy sites, which provides the endoscopist with a confident 3D region around the endoscope camera site and a measure of the reconstruction precision. As the experimental result shows, the root mean square target registration error of IGGBS using merging model is 9.5 mm, which is close to the conventional biopsy tattooing method (about 1cm).

Compared with the traditional 2D electronic map navigation, the visual navigation has large amount of data. Especially when the viewpoint changes with high speed, navigation devices can’t display smoothly. Based on the above problem, this paper carried out the research from three aspects which could influence the speed of terrain visualization, that were the storage index of terrain data, data scheduling and occlusion culling. By adopting appropriate storage method, putting forward a new data scheduling strategy, and improving occlusion culling algorithm, the efficiency of the terrain rendering had been improved effectively. Experiments show that this method can meet the basic requirements of visual navigation.

Recently, interactive character animations in computer games are mainly rely on motion-captured or carefully crafted motion clips. However, it is impractical and difficult to provide motion data samples for all possible behaviors with creating realistic responses to unexpected changes in the world. In order to control characters animations more precisely and realistically in real-time, an decision mechanism that synthesizes animations with interaction between characters in virtual environment is necessary. An efficient near-optimal interactive motion controller that can intelligently adopt the input motion data to a dynamically changing virtual environment is present. The controllers using reinforcement learning to reduce the data requirement by finding the most effective set of motion data to create the desired behaviors, and can be used as basic components in global path planning, and it further reduces the computational processing burden in the real-time interactive applications.

Melanin is one of the most important pigments in human skin. Estimating melanin distribution is required in many different research fields such as anthropology, dermatology and cosmetics. The current noninvasive estimation methods are mostly based on spectroscopy, which may not be available in many applications. In this paper, a fast estimation algorithm is proposed. Based on the principles of optics and skin biophysics, the process of the skin colour formation is inversed and modeled by an Elman network, and the corresponding melanin distribution can be obtained. The algorithm was evaluated on many skin images with different camera models and illumination conditions. Experimental results demonstrate that the proposed algorithm performs significantly better than other estimation methods. It can help researchers easily detect skin regions with abnormal amount of melanin pigment.

The target detection of digital image is one of the main content in computer vision research, which has a wider use. This paper presents an algorithm of the fuzzy small target detection for digital image. First, all the pixel values are looked as a set of elements with the corresponding address, and the small target is determined according to the need, so the image pixels are divided into two sets which includes target set and its complementary set; then the addresses of the storage target pixels are located; the next step to do is calculating the thresholds of target set and its complementary set; Finally, the binarization operation is applied to the small target set and its complement set by the calculated threshold. The test results show that this algorithm for small target detection is very effective.

We proposed a leaf vein visualization modeling method through deformation. Based on blade modeling which is on the basis of deformed rectangle is applied to the midrib and lateral vein which are made by conical function. In this way, different shapes of leaf veins produced are similar to what the realistic veins. The method is suitable for the generation of the midrib (principal vein) and lateral vein (secondary vein) of leaf, especially for the sleek lateral vein.

The motivation of the paper is to position and track moving targets in real time by means of wireless video sensor networks (WVSNs), and with the aid of selecting and associating information from multi-view videos to obtain the trajectory of targets in the world coordinates system(WCS). The whole process includes camera calibration, target localization, tracking trajectory in WCS, time synchronization calibration of multi-view videos and data fusion. The innovation point of the paper is that aimed at time synchronization phenomenon of multi-view videos while associating data, we propose a new scheme of information selection and recognition. For validating the effectiveness of the scheme, the comparison of the real trajectory and the estimated one is conducted through MATLAB simulation, and the proposed scheme has a satisfactory performance.

SVM is a new general learning method based on the statistic learning system which can be used as an effective means to process small sample, nonlinear and high dimensional pattern recognition. This paper did research on the learning algorithm of support vector machine, extracted characteristic data of banknote which is on account of PCA according to the characteristics of the support vector machine (SVM), and proposed to put support vector machine (SVM) into banknotes denomination recognition by combining SMO training algorithm with one-to-many multi-value classification algorithm. Besides, this article used genetic algorithm in parameters optimization such as the punishment coefficient C of soft margin SVM and the width parameter of Gaussian kernel function. The ultimate purpose is to recognize the denomination of banknote efficiently and accurately. The experimental results verified that this kind of recognition method increases the recognition accuracy up to 90% or more.

We proposed a parameter model method for outline of plant leaf which combining the shape feature of the outline and planar rectangle. The proposed method using deformation function to define the outline of plant leaf and it included leaf apex, leaf base and leaf margin .The method combined these deformation functions to form different shapes of the outline. We give some examples of representative outline of plant leaf. The method provided recognition with high speed and efficiency according to the results.

Image segmentation is to segment images into subdomains with same intensity, texture or color. Texture is one of the most important features to images. Because of the complexity of texture, segmentation of texture image is especially difficult and it seriously restricts the development of image processing. In this paper, a nonlocal Mumford-Shah (NLMS) model is proposed to segment multiphase texture images. This proposed model uses nonlocal operators that are capable of handling texture information in the image. In order to segment different patterns of texture simultaneously, multiple region partition strategy which uses

n

label functions to segment

n

+1 texture regions is adopted. Furthermore, to improve computational efficiency, the proposed model avoids directly computing the resulting nonlinear partial differential equation (PDE) by using Split Bregman algorithm. Numerical experiments are conducted to validate the performance of proposed model.

In this paper, an automatic and efficient cloud detection algorithm for multi-spectral high spatial resolution images is proposed. Based on the statistical properties and spectral properties on a large number of the imagery with cloud layers, a multispectral-based progressive optimal scheme for detecting clouds in Landsat-8 imagery is presented. First, a basic process which distinguishes the difference between cloud regions and non-cloud regions is constructed. Based on the spectral properties of cloud and the optimal threshold setting, we obtain a basic cloud detection result which separates the input imagery into the potential cloud pixels and non-cloud pixels. Then, the potential cloud regions and the cloud optimal map are used together to derive the potential cloud layer. An optimal process of probability for clouds over land and water is implemented with a combination of a normalized snow/ice inspection and spectral variability inspection. Finally, in order to obtain the accurate cloud regions from the potential cloud regions, a robust refinement process derived from a guided filter is constructed to guide us in removing non-cloud regions from the potential cloud regions. The boundaries of cloud regions and semi-transparent cloud regions are further marked to achieve the final cloud detection results. The proposed algorithm is implemented on the Landsat-8 imagery and evaluated in visual comparison and quantitative evaluation, and the cloud-covered regions were effectively detected without manual intervention.

This paper presents a novel approach for combining and blending a set of aligned images into a composite mosaic with no visible seams and minimal texture distortion. To accelerate execution speed in building high resolution mosaics, the compositing seam is found efficiently via seeded region growing using a photometric criterion. A contribution of this paper is to use seeded region growing on image differences to find possible seams over areas of low photometric difference. This can result in significant reduction of blending time. The proposed method presents several advantages. The using of seeded region growing over image pairs guarantees the approximate optimal solution for each intersection region. The independence of such regions makes the algorithm suitable for parallel implementation. The using of priority queue leads to reduced memory requirements and a compact storage of the input data. Finally, it allows the efficient creation of large mosaics, without user intervention. We also evaluate the proposed blending method with pixel based graph cut and watershed based graph cut to illustrate the performance of the approach on image sequences with qualitative and quantitative comparison.

The Fourier transform of the promotion is considered, which aims to develop a new model to solve complex data processing problems. This paper introduces the definition of multi-dimensional vector matrix. Based on the multi-dimensional vector matrix theory, the Fourier transformation is extended to multi-dimensional space, which includes the deduction of unitary orthogonal conjugate and energy concentration. Then the translation theory of two-dimension Fourier transform is extended to multi-dimension space.

In this paper, a small target detection algorithm in infrared image is proposed. First, an infrared image is coarse-to-fine segmented automatically by self-adaptive histogram segmentation. After detecting small abnormal region in segmented image and defining them to candidate targets, the abnormality based confidence of each candidate target is calculated and sorted. Finally, the candidate target with the maximum confidence is weighed to be real one. The experiments demonstrate that the proposed method is efficient and robust.

Based on the special geometry of cave targets, a method for LiDAR point cloud data rapid recognition of residents cave and natural cave has been proposed. By calculating the distance between LiDAR points and sensor position, the target can be quickly identified if there is a distance sudden change. In order to improve the accuracy of target recognition, the number threshold of sudden change points and region-wide threshold are introduced, and the recognized target contour points can be further distinguished in target front plane or inside plane.

In this paper, we propose a robust visual tracking method by exploiting both the structural and the context information. Firstly we take use of the sparse coding’s robust to occlusion and illumination and extract the structural local sparse feature, upon which we create a discriminative model between the target and the context. Then we introduce an adaptive online SVM algorithm to searching the feature space and discriminate the target from the context patches. Furthermore, the update of the dictionary and the SVM model consider both the latest observations and the original template, thereby enabling the tracker to deal with appearance change and alleviate the drift problem. Experiments compared with the state of art algorithm demonstrate that the proposed tracker performs excellent in the challenging videos.

In order to correct the distortion of C-arm projection image, a moving least squares combined with polynomial fitting global correction method was proposed. The pincushion distortion, S distortion and local distortion were considered as a whole and moving least squares was used to increase virtual control points for its compact support. The composite distortion was fitted by Nth polynomial and the images were corrected with optimum correction coefficients obtained by the least square method. The result demonstrates that the proposed method could achieve higher precision with less control points than only with global polynomial fitting, and the actual space error is 0.1054±0.068mm which means it is possible to apply the corrected images in surgery navigation.

Erratum to: Chapter 8: T. Tan et al. (Eds.)Advances in Image and Graphics TechnologiesDOI: 10.1007/978-3-662-47791-5_8