Research On Unsupervised Segmentation Method Of Color-texture Image Based On Multi-scale Structure Tensor

Image segmentation is a fundamental task in field of digital image processing andvision computing, and it can divide the original image into several non-overlapsub-regions and enclosed smoothly curves so that these segment results contain somespecial meanings. Consequently, image segmentation is becoming a basis topic forresearching. As we known, image segmentation is widely applied in object recognition,scene analysis, special effect in movie, object detection, medical image processing,industrial detection, and content based image retrieval and so on, the reason is that it canprovide some vital important cues for high level applications such as tracking, detectingand recognition et.al. However, as we lack the deeply understand about the visionmechanism of human, and the content contained in color-texture is diversity, complexity,and randomly. Thus, there is still not appeared a mature approach to meet with any kindsof need in application. Base on the actual circumstance, image segmentation is becominga difficult spot and hot topic, and how to improve the segmentation result to achieve thegeneralized and unified task, is still a challenge and arduous mission, and not tackledidealy until now.Aim at the imitated function of computer, such as micro structure, macro structure,and abstract information captured adaptively, and meanwhile it can further acquires theimportant semantic information. As a result, these functions lead to the unsupervisedsegmentation approach is focusing broadly, and it should achieve the segmentation tasks,such as distinguish and discriminate the different object accurately, recently. However,due to the abundant color information such as chrome，saturation，illumination andbrightness are contained in natural image. Along with the diversity attributes, such aslinearity, homogeneity, regularity, randomly, smoothly, roughness, density, granularity,frequency, phase, and directivity of texture, which are related to circumstancerelationship closely. Thus, the integrated method of color information and texture information is focused significantly in our article. Due to the requirement aboutrobustness, vigorously, and instantaneity in application, the research about Graph Cutsoptimization methods has become one of the most popular methods in recent years. It hasmany excellent features, such as with the ability to fuse a wide range of visual cues andconstraints, global optima, numerical robustness, practically efficiency, and unrestrictedtopological properties of weighted graph for segments etc. Nevertheless, the traditionalGraph Cut methods are not skilled in applicability, accuracy, robustness, and thereal-time, therefore, these factors are severely limited the usability of the Graph Cutmethods when applying in color texture image segmentation, such as complex, various,random, and class number uncertain situation. Therefore, to alleviate the key problemsmentioned above, we carry out a qualitative analysis and quantitative research,theoretical and practical research accordingly, then, several new unsupervisedcolor-texture segmentation approaches are proposed. Concretely, the main innovation andresearch achievements of this dissertation can be described as follows:Firstly, a new color-texture descriptor was constructed by integrating the compactmulti-scale structure tensor (MSST) texture, RGB color information, and the totalvariation flow (TV). As the MSST with the ability to compact the whole orientationinformation, multi-scale information description, and organic combination of frequencydomain and spatial information, therefore, we adopt it to extract the multi-scale textureinformation. But, for MSST, it is a set of matrix that meets with the nature of Riemannianmanifold, therefore, it will cost a great time consuming and memory burden whencalculating the geodesic distance and statistic value by using space mapping fromRiemannian manifold to tangent vector. More seriously, MMST is hard to integrate withcolor information directivity. To overcome these problems mentioned above, we use theSVD decomposition for each scale of MSST in tensor space, then, PCAalgorithm used toextract the main texture information. Due to the fact that MSST does not work well inlarge-scale texture region, then, the total variation flow is used to auxiliary describe the texture feature by extracting local scale information. Meanwhile, to suppress the noises,enhance the boundaries and improve the homogeneous of similar color texture objects,the nonlinear diffusion filtering is adopted.Secondly, a component-wise expectation-maximization for MMST algorithm(CEM~3ST) is proposed, it can calculate the valid class number through delete invalidcomponent when using sample support degree of each component of mixed MMSTdistribution. To simulate and speed up the information propagation of valid class, wereplace explicit scheme by semi-implicit scheme to update only one component at a time,and it can speed up the convergence. To exclude the impact of invalid component duringthe EM iterative procedure, the support degree of current component was calculatedaccording to all the samples, using the treated and non-treated valid components. Andthen we can delete the invalid component after the sample support degree normalized. Inother word, the invalid components can be quickly redistributed to the other validcomponents through recalculating the statistic value，valid class number，and samplesupport degree by adopting the least square and maximum likelihood (ML) process.Thirdly, we propose an edge-based and region-based multiphase successive activecontour model (MSACM). In details, we improve the constraints about the externalregion force and internal boundary force of Chan and Vese (CV) model; it can enhancethe homogeneous detection ability and noise tolerance through using the multivariableGMM model to describe probability density function (PDF). More important, we canbreak down the assumption about constant PDF in internal region. Except that, weincorporate geodesic active contour (GAC) into MSACM model to enhance the detectionability for concave edge and contrast with noise. As we knowledge, the optimal solutionof MSACM model is equivalent to energy minimal, usually, we can resolve the optimalproblem in numerical mode by using level set. Nevertheless, we confront with difficult,as the level set method is easily draped in local minimization and convergence slowly. Asa result, we resort to Cauchy-Crofton formula; then, the energy minimization problem will be converted as maximum flow/minimum cut problem of multi-layer graph cut(MLG) model. So that we can quickly get the approximate optimal solution in global,using Graph Cut. It’s noteworthy that our proposed MSACM model, it cannot onlycaptures the smooth and deep concave boundary, but also it can keep the segmentedregions are more integrated finally.Lastly, an unsupervised color-texture image segmentation method is proposed usingmultivariate mixed student’s t-distribution (MMST) and regional credibility merging(RCM) strategy. To overcome the probability density function (PDF) appears theGaussian, non-Gaussian, and heavy tailed distribution, we adopt the multivariate mixedstudent’s t-distribution to build up the PDF model. As it can integrate with CEM~3STalgorithm for getting valid class number, thus, it can further reduce the layer number ofMLG model greatly. Meanwhile, to overcome the segmented regions appearover-segmentation and error-segmentation phenomenon; a strategy of regional credibilitymerging (RCM) is presented by integrating the regional adjacency relationship, regionsize, common edge between regions, and regional color-texture dissimilarity-Jdivergence distance to calculate the RCM value for each two adjacent regions. Ifacquired RCM value is lower, the two adjacent regions will be merged or deleted.Additionally, in order to terminate the whole segmentation process adaptively, anadaptive iteration convergence criterion is designed, which combines the negativelogarithm of probability of all color-texture features with the Kullback-Leibler (KL)divergence for MMST. So that the segmented regions with outperforming visual entiretyand region consistency.