Similarity Matching in Computer Vision and Multimedia

Comparing two images, or an image and a model, is the fundamental operation for any retrieval systems. The similarity matching of two images can reside in the hierarchical levels from pixel-by-pixel level, feature space level, object level , and semantic level. In most systems of interest, a simple pixel-by-pixel comparison will not suffice: the difference that we determine must bear some correlation with the perceptual difference of the two images or with the difference between two adequate semantics associated to the two images. Similarity matching techniques are developed mostly for recognition of objects under several conditions of the distortion while similarity measures, on the other hand, are used in applications like image databases. Matching and dissimilar-ity measurement are not seldom based on the same techniques , but they differ in emphasis and applications. Similarity judgments play a central role in theories of human knowledge representation, behavior, and problem solving and as such they are considered to be a valuable tool in the study of human perception and cognition. Tver-sky [12] describes the similarity concept as ''an organizing principle by which individuals classify objects, form concepts , and make generalizations. " Retrieval by similarity, i.e. retrieving images which are similar to an already retrieved image (retrieval by example) or to a model or schema, is a relatively old idea. From the start it was clear that retrieval by similarity called for specific definitions of what it means to be similar. Smeulders et al. [11] discuss several types of similarity that need to be considered when one is analyzing a pair of images: similarity between features, similarity of object silhouettes, similarity of structural features, similarity of salient features , and similarity at the semantic level. Summarizing their ideas (see also the survey of Datta, et al [4]) one can identify three components that typically every system for retrieval by similarity needs to have: Extraction of features or image signatures from the images, and an efficient representation and storage strategy for this precomputed data. A set of similarity measures, each of which captures some perceptively meaningful definition of similarity, and which should be efficiently computable when matching an example with the whole database. A user interface for the choice of which definition of similarity should be applied for retrieval, presentation of retrieved images, and for supporting relevance feedback. Gudivada [9] has listed different possible types of similarity for retrieval: color similarity, texture similarity, …