A Pyramidal Approach to Content-Based Image Retrieval

Search based on image contents is an important issue in large image and video databases. In this paper, two methods for a better content-based image retrieval (CBIR) are presented, namely, the use of recognition kernel and locales. Features of model objects are extracted at levels that are most appropriate to yield only the necessary yet sufficient details, together they form the kernel. Instead of relying on image segmentation, a method of feature localization based on locales is developed. It is shown that the deployment of the recognition kernel and locales in a pyramidal (multiresolution) framework delivers good retrieval results.