Finding a small number of regions in an image using low-level features

Many computer vision applications, such as object recognition, active vision, and content-based image retrieval (CBIR) could be made both more efficient and effective if the objects of most interest could be segmented easily from the background. In this paper, we discuss how to compute and process low-level features in order to obtain "reasonable" regions for the putative objects. This process is a precursor to the detection of salient objects in an image, a subject discussed in a companion report [1]. Although considerable work has been done on image segmentation, there still does not exist an "off-the-shelf" solution applicable to all types of images. A major issue has been the lack of a good measure of quality of a particular segmentation. In this paper, three different measures are considered: the non-parametric measure (NP) proposed by Pauwels and Frederix [2], the modified Hubert index (MH) [3], and a threshold-based measure with a manually selected threshold. From the experimental results, we have found that the simple threshold-based measure gave consistently better results than the other two more complex, statistically based measures. The particular image segmentation method we have employed in this study is also described in detail. (C) 2002 Pattern Recognition Society. Published by Elsevier Science Ltd. All rights reserved.