Image Segmentation Based On Weighted Signed Graph Clustering And Markov Random Fields

Image segmentation refers to the process to divide an image into non-overlapping regions,according to similarity criteria of features or feature sets such as color,gray value,texture,and shape of the image.Image segmentation is one of the basic and core issues in computer vision,in which the unsigned graph clustering methods,such as Normalized cut(Ncut in short)and others,have successfully applications.Weighted signed graphs are able to express more relationships than unsigned graphs.As the most graph clustering methods are commonly based on unsigned graphs,this paper attempts to build a semi-supervised image segmentation based on weighted signed graph clustering,analyzes Signed Normalized cut(Signed Ncut in short)on weighted signed graphs,and propose improvements based on the MRF regularization.The work of this paper is as followed:(1)This paper expresses semi-supervised information as pairwise constraints,and then embeds them into a weighted signed graph on which Signed Normalized cut is used to clustering analyses.Spectral method,a main optimization method of the graph clustering,is adopted to solve the relaxation problem,and the k-means algorithm is adopted to do image binary segmentation.The result shows that Signed Ncut performs better than Ncut in overall situation,which suggests Signed Ncut is viable in binary segmentation.However,Signed Ncut needs further implements to improve segmentation accuracy,other statistical indicators,and contour fit.(2)A regularized potential function is fit into the MRF regularization,to improve the segmentation utility of Signed Ncut.Based on Graph cuts technology,the research applies the linear approximation to Signed Ncut,constructs an upper bound auxiliary function and approximate the optimal solution iteratively.The results show that MRF regularized Signed Ncut perform better in contour fit and related statistical indicators.In summary,this paper attempts to use the weighted signed graph to express the relationship between the pixels,fuses Signed Ncut and MRF regularization,provides a new method for image segmentation,and also enlarges the application of signed graph clustering.