Image retrieval at low bit rates: BSP Trees vs. JPEG

Content based image retrieval (CBIR) has still a long way to go before it will become capable of distinguishing similar rather than almost identical images. These limitations result both from a lack of ‘intelligent ’ enough algorithms as well as problems in terms of computational restrictions which force algorithms to be as simple as possible, while maintaining reasonably high effectiveness. Images are not analysed in terms of their physical or logical contents in the way humans perceive it but most often as statistical information such as spatial correlation of colors or intensities of separate pixels. This information is usually extracted at a very local (pixel) level and may be miss-interpreted if pixel values change (even though the overall image will still appear similar), which is especially the case in image compression. Results of earlier experiments [10] show that this can cause notable problems for CBIR algorithms. In this paper, we try to address the problem by investigating a more appropriate compression algorithm based on binary space partitioning trees and how it can improve the retrieval performance of compressed images. Simulations on two different image databases show that BSP compression outperforms JPEG for colour-based image retrieval while for texture-based indexing JPEG performs better