Research On Random Walks Image Segmentation Method

According to certain similar criteria of some low-level visual features, such as color, texture, shape etc, image segmentation refers to partition a single image into non-overlapping homogeneous regions, and then extracting the objects of interest. Image segmentation is the basis of high-level applications, such as feature extraction and object recognition et.al. Consequently, it has been used in many fields, such as communication, military, remote sensing, meteorology, medical image processing, intelligent transportation and automated detection of agricultural and industrial. However, as the content contained in nature color image is diversity, complexity and randomly, and the "sickness" of the segmentation method, together that we lack the deeply understand about the vision nechanism of human. There has not been a very mature segmentation method can meet the various needs of different application environments. Therefore, to further study the segmentation approach and design a method with accurate, efficient, universal and practical will be important significance to the development of image processing and computer vision.Due to the lack of prior information about the objects in an image, Automatic approach is hard to provide the ideal segmentation results about real-world natural scene images. Interactive image segmentation methods incorporate minimal user interactive into the segmentation process. Due to its good segmentation performance, interactive method has attracted significant attentions in recent years. However, the segmentation quality depends heavily on the priori information provided by user and the manual interactions are time-consuming, it all limits the usability of current interactive tools for image segmentation. To address the above problems, focusing on interactive image segmentation method based on graph theory, this work mainly studies the automatic methods of combining saliency estimate method with interactive random walks segmentation methods. Additionly, an effective region saliency estimate method based on entropy rate superpixel segmentation method is proposed. By combining the multiscale nonlinear structure tensor matrix theory, dimension reduction and manifold learning method, new feature modeling methods used for improve segmentation results were proposed in this dissertation. It has achieved satisfying results when used in automatic image segmentation and interactive multi-class object segmentation algorithm. The main innovative research achievements of this dissertation can be described as follows.Firstly, to address the problem of the existing saliency estimate methods, an effective region saliency estimate method based on entropy rate superpixel segmentation method (region saliency based on entropy rate superpixel, RSBERS) is proposed. In order to preserve the local structure of image and remove the unnecessary details, we first apply entropy rate superpixel method to segment the input image into superpixel regions. And then compute saliency value for each region to obtain the global visual saliency map. RSBERS can highlight the whole salient objects uniformly. Meanwhile, it has the ability to well-defined boundaries of the salient object.Secondly, this work proposes a novel automatic image segmentation method by combining interactive random walks method with an automated seeds extraction method. In order to extract seeds with sufficient and accurate information automatically, we apply the affinity propagation to each superpixel region to get representative pixels, which are then labeled by a fixed threshold to obtain seeds. Meanwhile, a seeds relabeling method is presented and can be used to reduce the mislabeled seeds. Finally, as heuristic information, we use random walks method to obtain segmentation results. In addition, to improve the powerful feature discriminating ability, a new feature descriptor is designed by using covariance matrices of low-level features, contationing coordinates information, L*a*b*color information and multi-scale nonlinear structure tensor texture information, for local neighborhoods.Finally, as the multiscale nonlinear structure tensor with the ability to compact the whole orientation information, and also multiscale information description, we adopt it to extract the texture information. However, we split the multiscale nonlinear structure tensor into a series of independent component when integrate with other information, it destroy the integrity of matrix. To overcome these problems mentioned above, we use Isomap algorithm for each scale of multiscale nonlinear structure tensor in tensor space to obtain compact texture information.