Color Image Segmentation Based on Regional Saliency

Abstract

In this paper, we propose a novel segmentation model integrated the salient regional features into mean shift (MS) clustering segmentation as fusion matrixes. Firstly, a regional visual saliency map of the given image is obtained based on quantification image in HSV color space. Then saliency factors are extracted from salience map from each channel in L*a*b space in two steps: region saliency(S-R) and pixels-region (P-R). Fuse the salient factors derived from former salient features with original components of the image as new input features, who are involved in the mean-shift procedure for segmentation. This paper takes advantage of regional salience to guide the MS vectors moving to accurate modes, and decreases premature and ill convergence at local area. The introduction of salient factors enhances the accuracy of the pixels clustering for region segment. Experiment results carried on Berkeley database and comparison with human segmentation results demonstrated that our algorithm has better performance on nature color images segmentation.