Content-Based Image Retrieval Using Regional Representation.

Representing general images using global features extracted from the entire image may be inappropriate because the images often contain several objects or regions that are totally different from each other in terms of visual image properties. These features cannot adequately represent the variations and hence fail to describe the image content correctly. We advocate the use of features extracted from image regions and represent the images by a set of regional features. In our work, an image is segmented into "homogeneous" regions using a histogram clustering algorithm. Each image is then represented by a set of regions with region descriptors. Region descriptors consist of feature vectors representing color, texture, area and location of regions. Image similarity is measured by a newly proposed Region Match Distance metric for comparing images by region similarity. Comparison of image retrieval using global and regional features is presented and the advantage of using regional representation is demonstrated.