Color image adaptive segmentation based on rival penalized competitive learning

Abstract

For color image clustering segmentation, selection of the number of its main colors is crucial. Considering the fact that rival penalized competitive learning (RPCL) can converge weights of some initial units to actual centers of input dataset's classes if the number of initial units is larger than that of dataset's classes, and this paper proposes an adaptive segmentation approach for color image in RGB space based on RPCL. Firstly, histograms of R, G, B component colors are segmented by adaptive threshold segmentation algorithm for gray image. Secondly, some possible colors of original image are specified through combining those components' segmented areas and false colors not appearing in image are removed. Finally those possible colors are converged to the main actual colors of original image by RPCL. Then original image is segmented by those learned centers. In the end the result of experiment on mobile robot CASIA-1 shows that the approach can adaptively segment a color image into several classes without specifying the number of initial classes in advance.