Research On Image Segmentation Based On Graph Cuts And Shape Prior Constraints

The task of image segmentation is to seperate one image into several nonoverlapping regions and extract a region of interest based on certain characteristics(such as color, gray scale, texture, shape, etc.), and these characteristics in the same area is similarity, show obvious differences in different areas. Over the years, although researchers have proposed many algorithms, but so far neither general segmentation algorithms, nor objective evaluation criteria of segmentation result. Due to the current image segmentation technology has not yet reached the level of understanding image, sometimes automatic segmentation algorithm is difficult to achieve good results, so interactive segmentation has received more and more researchers’ attention.Before an image is divided into object and background, certain hard constraints from user will be imposed, we indicate certain pixels to be part of the object and certain pixels to be part of the background, which make the pixels with ambiguity classification more accurate. In numerous interactive segmentation algorithms, because of fast and global optimality, Graph Cuts algorithm gets a lot of researchers’ attention. By introducing graph theory into image segmentation, Graph Cuts make use of the mapping relation between the pixels and the nodes in graph to build the graph model, which can transform the problem of global optimum to min cut of graph.For the weak boundary, complex background and noise, a priori shape model can be used as constraint to reduce the ambiguity of segmentation, and improve the quality of image segmentation.We consider the image segmentation based on the Graph Cuts algorithm and the prior shape model, including:1. Improving Graph Cuts model by using the distance metric to replace the Euclidean distance. To learn metric matrix, the pixels are divided into three parts: similar sample set with labels, dissimilar sample set with labels, set without labels,one principle for metric learning is to minimize the distances between the data points in similar set and to maximize the distances between the data points in dissimilar set.2. In order to overcome absence of the effective object and background model, based on simple geometric property of object, we propose image segmentation based on ray shape priors. This algorithm needs to mark the center of object, according to the location relation between two adjacent pixels and the center, we set weights in advance. In order to make the algorithm more applicable, we replace one center by multiple centers.3. For the complex background, we propose image segmentation based on sparse shape priors. KPCA is used to learn shape priors feature space, the object shape probability that is the linear combination of prior shapes and the coefficients are sparse is estimated in the feature space, then the object shape probability as constraint is combined with Graph Cuts.