Image Segmentation Method Based On Super - Pixel Spectrum Clustering

Image segmentation is the process of dividing an image into a plurality of communicating with consistency and similarity in the visual area. As a critical step in image analysis and pattern recognition,the image segmentation has been a hot topic in computer vision. Researchers have proposed many techniques for image segmentation algorithm, but so far still no one is appropriate for all images. Compared with the conventional image segmentation algorithm, in recent years the segmentation method based on spectral clustering becoming a hot topic has shown a clear superiority in a variety of experiments and simulations. Therefore, this paper is based on the normalized cut and the concept of superpixel, then builds an improved similarity matrix combined with superpixel color information space and uses the normalized similarity matrix segmentation method for clustering segmentation. The main contents are as follows:Firstly, the paper introduces the basic concepts of image segmentation and the development of image segmentation. We introduces several of the important image segmentation methods, including image segmentation method based on threshold, image segmentation method based on the edge of the area and image segmentation methods based on the specific theory.Secondly, the paper introduces the relevant basic theoryof graph, two kinds of image segmentation criteria which based on graph theory, and solution of nomalized cut algorithm. Then this chapter introduced the2-way Ncut and K-way Ncut algorithm. Finally, simulation results show that the select of parameters K and the weight matrix W in the K-way Ncut algorithm is very important. Furthermore, we introduces the principle of mean shift algorithm, including its basic form, expanded form, steps of the algorithm, and discuss the convergence of the algorithm, analysis the image segmentation based on mean shift algorithm. Simulation results show that the selection of the three parameters (hr,hs,M) in the algorithm has great impact on segmentation results.Finally,the paper proposes the segmentation algorithm of the spectral clustering based on super pixel combined the normalized segmentation algorithm theory with mean shift algorithm. The algorithm use mean shift algorithim to preprocess the image, as the image processed by mean shift algorithm will be divided into many over-segmentation superpixels regions, we use some key point instead of these superpixels regions to construct an undirected weighted graph to describe the relationship of these point, unlike the traditional Ncut algorithm used the image pixels. Experimental results demonstrate that the proposed method can not only improve the segmentation of the image, but also reduce the computational complexity for real-time image processing.