A Fuzzy Membership Model for FSVR-Based Image Coding

Abstract

In this paper, a modeling method of fuzzy membership based on data domain description is proposed for image coding by fuzzy support vector regression. The original image is divided into some non-overlapped rectangular blocks and their transform domain coefficients are treated as training data sets. On each data set, data points are nonlinearly mapped into a high dimensional feature space where the smallest enclosing hypersphere is obtained. Then the corresponding fuzzy membership model is constructed from the distance of each point to the center of the sphere. The established model is eventually embedded into the image coding scheme which adopts adaptively variable penalty factors. Experimental results show that the proposed approach achieves improved quality in both subjective and objective measurement.