Research Of Image Segmentation Based On Sparse Representation

Image segmentation is a fundamental low-level vision problem with the purpose of segmenting the object from the background. It has been used in many computer vision applications. With the developing of some subjects, many image segmentation algorithms have been proposed. Image segmentation based on graph theory mapped image into a weighted undirected graph and combined the graph theory to segmenting images. Based on the bipartite graph (Segmentation by Aggregating Superpixels, SAS), image segmentation algorithm can simultaneously consider the space organizational relationships between the superpixels and pixels as well as the different superpixels.It was a problem about how to segmenting image correctly. With the rapid developing of sparse representation (SR), the segmentation algorithm based on sparse representation was proposed. Introducing of l0 sparse representation has the advance of keeping the global features and semantic segmentation results. However this method connects superpixels with the superpixels which can reconstruct images, but it ignores key links between adjacent superpixels. In addition, sparse representation method greatly increases the calculation complexity of the method.In this paper, we proposed a bipartite graph model based on collaborative representation. The experimental results show that the proposed model has good performance with image segmentation. Our main works are summarized as follows:Firstly, we analysis the sparse representation image segmentation algorithm based on compressed sensing.Secondly, based on the SAS method, we reconstruct the bipartite graph using Lab color space and local binary pattern (LBP).Thirdly, to the purpose of reducing the time of l0-SAS method, we combined the collaborative representation (CR) method to constructing a new image segmentation algorithm. This algorithm can ensure the effectiveness of the algorithm and greatly reduce the complexity.Finally, a series of simulation experiments on image segmentation databases of BSD300 and ore image segmentation have been done to prove the accuracy of the CR-SAS algorithm.