Project Scientific Results

Project title: Lossy image coding based on invariant features across scales and orientations, and its evaluation using a novel definition of the best achievable compression ratio in lossy coding

Supported by: CICYT (Spain). Project number: TIC2000-1421

Starting and ending dates: from December 2000 to December 2003

Members of the team: J. A. Garcia (Head) and M. C. Aranda and J. Fdez-Valdivia and Xose R. Fdez-Vidal and J. Martinez Baena and R Rodriguez Sanchez

Description: The RGFF representational model can be used to decompose the original scene into a number of images isolating statistical structures which maximize the redundancy across scales and orientations. The derived redundancies can then be exploted to decrease the number of bits that are needed to code the original scene. But the problem of lossy image coding techniques is that there is a trade-off between image distortion and coding rate. This trade-off may be reached with several techniques, but all of them require an ability to measure distortion. And finding a general enough measure of perceptual quality has proven to be an elusive goal. To circumvent the lack of knowledge of what distortion measures are more suitable for images, here we propose to develop a novel technique for deriving optimal performance bounds (it has been termed the "best achievable" compression ratio) based on the relationship between information theory and the problem of testing hypotheses. The best achievable compression ratio for lossy coders will determine a boundary between achievable and non-achievable regions in the trade-off between source fidelity and coding rate. The resultant bounds will be tight for situations of practical relevance (i.e., correspond to high coding rate). These performance bounds will be directly achievable by a constructive procedure as suggested in a theorem which is intended to prove the relationship between the "best achievable" compression ratio and the Kullback-Leibler information gain. We will test the best achievable compression ratio for various lossy coding schemes and several kind of scenes. We also evaluate the different coding techniques in the rate-distortion sense by using the concept of the best achievable compression ratio and the corresponding error probability in a bayesian setting. These results will provide an insight into the design issues of optimizing lossy coders, as well as a good reference for application developers to choose from an increasingly large family of lossy image coders for their applications