The Construction And Analysis Of Signed Networks For Image Segmentation

Signed network is a kind of network including edges with the property of positive or negative sign.The positive and negative sign represent positive relationship and negative relationship respectively.This negative relations commonly exist in fields of information,biology and society,and provides important application value for people's attitude prediction,user characteristics analysis and clustering and so on.In recent years,there are many signed graph clustering criteria,which have been successfully applied in many fields,and verified the added value of negative edge.However,how to understand the relationship and difference between signed graph clustering and typical unsigned graph clustering,how to unify multiple signed graph clustering,and how to construct symbol network for application scenarios,these problems have not been deeply studied.This paper focuses on the practical problem of image segmentation,and above questions are discussed based on the typical clustering criterion,Normalized cut(Ncut in short),which is successfully applied in this field.Image as a kind of semi-structured data,often use non negative weight to express the similarity between pixels(blocks)in traditional image clustering algorithm.In fact,the negative relationship also exists in the image data analysis.The pixel(block)pairs of different segmentation areas are mutually exclusive,which can be represented by negative edge.Compared with the unsigned graph,the signed graph with negative edge is more expressive.In this paper,signed network clustering is used to solve the problem of semi supervised image segmentation,and the construction of signed network is studied and the clustering of signed graph is analyzed in detail.The main work of this paper includes the following aspects.1.Normalized cut on signed graphs are compared and analyzed.Several kinds of semi-supervised Normalized cut based on graph are compared and analyzed.First of all,the similarities and differencesbetween semi-supervised normalized cut on unsigned networks and signed graph normalized cut are analyzed,and their unified definitions are given.Then,we compare and analyze many kinds of signed normalized cuts,propose four goals of signed network clustering,and define generalized signed normalized cuts.Finally,the experimental analysis on the interactive image segmentation in the form of scribing is carried out,and the results show that: i)The definition of graph will have a significant effect on clustering performance,and the normalization form will also have a certain effect on the results.ii)In order to achieve the two goals of minimizing the intra group similarity and maximizing the intra group similarity,clustering can be carried out well,and only considering the inter group similarity can also provide effective guidance for clustering.2.Aiming at the image segmentation problem,signed networks are constructed in two methods of priori form and the effectiveness of the presented approaches is validated.Signed networks constructed by touch interaction and the saliency detection,and use kernel-cut algorithm to optimize signed normalized cut under MRF constraints.By conducting segmentation experiments on image data,we found that MRF can improve the segmentation effect.Two transcendental forms of constructing signed networks are active,and the results of interactive segmentation based are slightly better than automatic segmentation based on saliency detection.In conclusion,we attempt to construct signed networks for image data,and analyze network clustering criterions to complete the semi-supervised image segmentation.It not only provides a new method for semi-supervised clustering,but also expands the application of signed networks.