An overview and evaluation of JPEG compressed domain retrieval techniques

While content-based image retrieval (CBIR) has been an active research area over many years, most CBIR techniques operate in the pixel domain even though images are typically stored in compressed form. Consequently, image decoding is required prior to feature calculation leading to a computational overhead that is prohibitive in particular for the case of online retrieval. However, as has been shown by several authors, well performing image features can also be extracted directly from the compressed domain of the images. In this paper, we focus on JPEG compression since it represents the dominant image format, and provide an overview of JPEG compressed domain CBIR algorithms. We furthermore compare these algorithms on a common benchmark database, the MPEG-7 Common Colour dataset. We demonstrate conclusively that working in the compressed domain is consistently faster than in the pixel domain, typically requiring around 15% of the processing time. We also show that in several cases the retrieval performance is comparable to that of methods in the pixel domain whilst maintaining this speed gain.