Color image segmentation using rival penalized controlled competitive learning

Color image segmentation has been extensively applied to a lot of applications such as pattern recognition, image compression and matching. In the literature, conventional k-means is one common algorithm used in pixel-based image segmentation. However, it needs to pre-assign an appropriate cluster number before performing clustering, which is an intractable problem from a practical viewpoint. In contrast, the recently proposed rival penalization controlled competitive learning (RPCCL) approach (Cheung, 2002) can perform correct clustering without knowing the exact cluster number in analog with the RPCL (Xu et al., 1993). The RPCCL penalizes the rivals with a strength control such that extra seed points are automatically driven far away from the input data set, but without the de-learning rate selecting problem as the RPCL. In this paper, we further investigate the RPCCL on color image segmentation in comparison with the k-means and RPCL algorithms.

I. Introduction

Color image segmentation has been extensively applied in pattern recognition, image analysis and computer vision such as image retrieval [1], face recognition and image compression [6]. In the literature, one image segmentation approach is pixel-based segmentation, in which the conventional means clustering algorithm [3] is commonly used to classify color pixels into different clusters [5], [7]. The -means generally requests to pre-assign an appropriate cluster number . Otherwise, its performance may seriously deteriorate. Unfortunately, such a value selection is an intractable problem from a practical viewpoint.