Novel multiresolution metrics for content-based image retrieval

This paper proposes three new ideas and one revision about image metrics and their applications. Of most importance is the multiresolution `shape metric', which measures distances according to images' content shape information. Accompanied with it is a feature/non-feature image characterization philosophy. The second effort is the modification of an existing `color metric', whose distance computation is dominated by color distribution of images. The third idea is on analyzing image querying behavior of the general public, so as to propose and design a versatile `power metric', which holds the properties of both the `shape metric' and the `color metric', for dealing with various kinds of hand-drawn queries. Finally, after applying the `shape metric' to video shot boundary detection, a multiresolution scene change detection algorithm is proposed. For now, we have implemented these metrics on a personal computer for two applications, one is for image database management with content-based indexing and the other is for video shot boundary detection. Experiments show that the speed and the accuracy of both image retrieval and scene change detection are quite well.