Study On Image Segmentation Methods Based On Sparse Regularization And Clustering

Image segmentation is one of the hot and difficult issues in the field of computer vision and image analysis.The goal of the image segmentation is to partition a given image into several non-overlapped regions according to some similarity characteristics(texture,intensity and color,etc.).Images are different,so there are many kinds of image segmentation methods.Recently,two-stage image segmentation methods have been widely paid attention by many researchers,due to their flexible structure and excellent performance.Generally speaking,the first stage is to find a smoothed image that can facilitate the segmentation,and the second stage is to segment the smoothed image.This paper proposes a two-stage segmentation method.The first stage proposes a novel discrete smoothing model for the images including texture,noise and blur.This model uses a sparse regularization(RTV,relative total variation)instead of the TV(Total variation)regularization in the model proposed by Cai.et al.in [SIAM J.Imaging Sciences,2013,6(1):368-390].This sparse regularization can highlight the edge of target in image very well.The second stage,the segmentation is done using the K-means clustering method,which has many advantages over other clustering segmentation methods.Because our model is not convex,we decompose the RTV measure into two nonlinear term and a quadratic term.The advantage is that the problem with the non-linear part can be transformed to solving a series of linear equation systems.Experimental results show the effectiveness of our method.Finally,we discuss the selection of the parameters for the proposed algorithm.