Query processing issues in image (multimedia) databases

Abstract

Multimedia database systems are essential for the effective and efficient use of large collections of image data. The aim of such systems is to enable retrieval of images based on their contents. As part of our research in this area, we are building a prototype content-based image retrieval system called CHITRA. This uses a four-level data model, and we have defined a fuzzy object query language (FOQL) for this system. This system enables retrieval based on high-level concepts, such as "retrieve images of mountains and sunset". A problem faced in this system is the processing of complex queries such as "retrieve all images that have a similar color histogram and a similar texture to the given example image". Such problems have attracted research attention in recent times. R. Fagin (1996) has given an algorithm for processing such queries and provided a probabilistic upper bound for the complexity of the algorithm (which has been implemented in IBM's Garlic project). In this paper, we provide a theoretical (probabilistic) analysis of the expected cost of this algorithm. We propose a new multi-step query processing algorithm and prove that it performs better than Fagin's algorithm in all cases. Our algorithm requires fewer database accesses. We have evaluated both algorithms against an image database of 1000 images on our CHITRA system. We have used both color histogram and Gabor texture features. Our analysis is presented and the reported experimental results validate our algorithm (which has a significant performance improvement).