Image segmentation via multi dimensional color transform and consensus based region merging

Abstract

Most natural scene real world images contain a lot of variations and produce a set of complex information. This complex information representation inducts difficulty in separating focal objects in the image. In this paper, hybrid cues are determined to efficiently separate the foreground objects from image background. The segmentation approach presented in our paper consists of two steps: 1) Production of clustering based super pixels 2) Consensus based optimal region merging process. First, input image is processed by mean shift as a noise removal step, then FCM clustering is employed to cluster pixels by utilising features based on extended color space transformations, following that final region labelling is done in terms of superpixel spatial connectivity. Second, hybrid cues are calculated as a tool for similarity measurement between regions, and a Consensus based region merging process is implemented by adjacent region similarity comparison with the standard deviation serving as a merging threshold, producing final segmentation. Experiments are conducted on Berkeley Segmentation Database and segmented images verify the efficiency of our approach in Natural Scene images.