Research On Image Segmentation Method Based On Probability Graph Model And Shape A Priori

Image segmentation is a very important research content in computer vision and image engineering. It is a key step from image processing to analyzing and understanding. Probabilistic graphical models based image segmentation methods provide a solid theoretical foundation for solving the problem of large amounts uncertainty in image information processing, and these models have a very sound theoretical framework, thus researchers continued to give in-depth research and quickly development in them. They have been widely used in computer vision, medical and remote sensing image processing and so on.Traditional image segmentation methods mainly rely on pixel color features, or adding some low-level features like edge and texture. These methods can get a good segmentation result when the object and background is smooth or their features easy to distinguish, but for the presence of occlusion, shadows, noise, clutter background or similar features it is often difficult to segment the object accurately and completely. To solve this problem, in this paper we introduce a high-level and global feature-shape into Markov random field and conditional random field in different ways, using shape prior constrained object boundary to guide a segmentation, we studied probabilistic graphical models and shape prior based image segmentation methods systematically.The main work and innovations are as follows:(1) This paper summarizes and analyses the research status and development of probabilistic graphical models and shape prior based image segmentation methods in detail. It provides an overview of probabilistic graphical models related theoretical basis, including their research and development history, concepts and types, commonly used Markov random field and conditional random field in image segmentation applications, their inference algorithms and parameter estimation.(2) A Markov random field based image segmentation associate with consistent constraint of shape prior and edge is proposed. This method combined shape prior of template with image edge features using Canny operator after their distance transform to get a more reliable consistent boundary constraint. Then using the constraint and color features to definite a segmentation energy function. The energy function is minimized by graph cuts to get the segmentation result. Experiments show that this method can obtain a more complete object when occlusion, shadows, clutter background, object boundary not clear or noise existing in an image compared to only using color feature algorithm and combining with shape prior algorithm. Furthermore, the consistent boundary constraint can reduce the shape differences impact after alignment a single shape template.(3) A conditional random field based image segmentation associate with shape prior is proposed. This method uses a discrete form of shape distance in level set to measure shape similarity and define the shape prior energy, through a shape alignment of shape template and object shape making the shape prior affine invariance. Using kernel method estimate the probability distribution of color features, which is more accurate and can reduce the number of parameters and its estimating time. This method gives the energy function definition of color and shape prior, and the spatial location of image pixels have contained in CRF, using graph cuts minimized the whole energy function to get the segmentation result. Experiments are given respectively with a single shape template to segment one object and multiple similar objects; they showed that introducing shape prior can effectively cope with the presence of noise, occlusion, complex background, lack of information, similar features, etc.