MPEG-7 Visual Descriptors and Discriminant Analysis

The MPEG-7 standards defines a set of descriptors to characterize the content of visual media [l, 2]. These visual descriptors, such as color and texture descriptors, have undergone extensive evaluation and development based on the application of retrieval ranking. Specifically, under query-by-example (QBE) paradigm, average normalized modified retrieval rank (ANMRR), a rate-accuracy like performance measure, is adopted to test these descriptors on image collection and predefined ground truth datasets. The experimental results show that each descriptor has good retrieval performance. However, there are some questions left to be answered in practice. How to apply visual descriptors in various applications? Does each visual descriptor have good performance in the applications besides retrieval ranking? How to combine multiple visual descriptors for a specific application? What is the performance of the aggregated visual descriptors? It would be generally accepted that a good visual descriptor should have excellent ability to separate distinct visual media content, named discriminant power. In various applications, the discriminant power of visual descriptors would be evaluated by the application-dependent performance criteria. Since the core experiments applied for the MPEG-7 standards concentrate on single visual descriptor and retrieval ranking, the discriminant power of visual descriptors has not been sufficiently evaluated. Particularly, the applications and technologies should be taken into account for evaluating the discriminant power of visual descriptors. This chapter answers the above questions from the perspective of discriminant power.