Content Based Approximate Picture Retrieval

In this paper real world pictures are represented by contents for storage and retrieval. The basic concept is that a picture has a foreground content and a background content, each of which may be represented by a collection of objects associated by some relationships. The E-R model is modiied to represent the pictures. We present sample mappings from restricted natural language expressions that correspond to user query descriptions into formal query language expressions. A knowledge base containing a thesaurus, generalization hierarchies and certain rules (permitting deduction) is used to enhance the performance. Pictures are not limited to a speciic domain. Similarity measures are calculated for complete and partial match conditions between the user's description and the pictures stored in the system. An elaborate description of the computation of the similarity is provided. Pictures which are suuciently similar to the user's description are retrieved. We present a sample retrieval as performed by our proyotype system. Preliminary experimental results are very encouraging.