University of Houston

Content based image retrieval involves the processes of querying and retrieving images from an image database using the image feature descriptors rather than its corresponding text annotations. Different descriptors such as MSER and DAISY have been shown to have different strengths in identifying images. We study some of these descriptors and using SURF and DAISY, we show that a combination of such descriptors perform better object recognition and classification than individual descriptors. Also, we concentrate on studying data structures that are most suitable to store and quickly retrieve such image features of very high dimensions. Since we are only concerned with the approximate nearest neighbor query problem, we implement the "Location Sensitive Hashing" ( LSH ) technique and the approximate kd-tree technique over SURF and DAISY descriptors combination and try to improve our query response time with very little loss in accuracy. We then show how our robust model can b e deployed on parallel and distributed environments, for batch processing of image queries.