Texture-based segmentation for extracting image shape features

Abstract

Shape features are one of the most popular low-level image representations for computer vision (CV) tasks such as template matching, image collaboration and object recognition. In this paper, an application-originated research has been introduced for extracting representative shape characteristics from challenging real-world scenes based on the image 'textures'. The proposed new approach starts from registering image colour and texture regions within undirected weight graphs. Then through applying Mean-Shift clustering, the graph can be used to identify image regions that contain similar texture patterns judging by the pair-wise region comparison operations. Based on the theoretical study and practical trials carried out in the research, the devised clustering-based segmentation strategy has proven its effectiveness under complex real-world conditions. The innovative I-PWRC algorithm developed in this research has integrated a number of the state-of-the-art image processing techniques including MS, PWRC, and the hierarchical pyramid structures. Test and evaluations have recorded satisfactory segmentation outputs and indicated its promising perspective for future CV applications, including video processing.