A watershed-based image segmentation using JND property

Image segmentation is the basic process in many image/video applications, such as computer vision, image analysis, medical imaging and the recent object oriented MPEG-4. Among proposed image segmentation algorithms, watershed is one of the most popular; however, the watershed algorithm suffers from an over-segmentation problem. Resolving the over-segmentation problem to obtain a concise region representation has been the focus of many researchers. We analyze and improve the preprocessing of the watershed algorithm and proceed to region merging using the human visual property of JND (just noticeable difference). Our goal is an image segmentation algorithm with the following three characteristics: (1) concise region representation which is consistent with human visual perception; (2) robust segmentation for a variety of image types; (3) efficient computation. We compare the proposed algorithm with two more sophisticated and computationally intensive segmentation algorithms; the results show that with the simple, yet very effective, JND merge criteria, the proposed algorithm is capable of generating region representations, which are concise and are more consistent with human visual perception for a varied spectrum of images.