Section: New Results
Joint source-channel and distributed coding
Joint source-channel coding based arithmetic codes with erasures
Participants : Simon Bos, Christine Guillemot, Simon Malinowski, Aline Roumy.
In the context of image or video transmission over noisy channels, the compression stage makes the transmitted message very sensitive to channel noise. Two kinds of codes are widely used in order to compress the signal : variable length and quasi-arithmetic codes. In the TEMICS project-team, a state model to be used in soft-decision (or trellis) decoding of variable length codes and quasi-arithmetic codes has been developed. This state model, compared to the optimal model, presents the advantage of achieving close to optimal decoding performance for a much lower complexity. This model is also well suited for the introduction and exploitation of a priori information (or side information) on the source in order to favor the selection of synchronous and correct paths in the soft decoding process. Two main joint source-channel coding strategies have thus been developed in the team. The first one consists in adding synchronisation markers at some instants of the decoding process. The second one is based on adding constraints on the number of decoded symbols at several instants of the decoding process. These approaches turn out to outperform widely used techniques, such as the popular approach based on the introduction of a forbidden symbol to quasi-arithmetic codes. These results have been obtained for AWGN channels, which represents transmission over mobile channels for example.
When dealing with image or video transmission over a packet based network (internet for instance), the issues are not the same. When a message (i.e., a bitstream) is sent over an Internet channel, it is subject to packet losses. This kind of channel can be modeled by an erasure channel (BEC) characterized by an probability of erasure. Techniques that are mainly used in practical schemes concerning the protection of video streams for transmission over packet-based network rely on forward error correcting (FEC) codes (Raptor codes, LDPC-Staircase...). Residual losses at the output of the FEC decoder then correspond to regions - which may be large - in images which have to be concealed. Concealment methods often make use of spatial and/or temporal interpolation techniques. In the context of the joint laboratory between INRIA and Alcatel Lucent, we are investigating methods of robust decoding of arithmetic codes in presence of erasures. The aim is to further improve the robustness of the video compression algorithms by confining the effects of losses to reduced areas, thus reducing the visual artefacts which result from concealment applied on large image areas. Methods of inpainting for recovering erased symbols are also investigated.
Slepian-Wolf coding of memory sources
Participants : Christine Guillemot, Aline Roumy, Velotiaray Toto-Zarasoa.
The problem of asymmetric and symmetric distributed coding for memoryless correlated sources has been addressed in 2008. Practical schemes based on LDPC codes have been designed. The minimum achievable rates depend on the correlation between the sources, but also on their respective distributions. Knowing the non-uniform source distribution, one can achieve lower rates using the same code. Syndrome-based turbo and LDPC decoders have been adapted so as to take into account the non-uniformity of the sources. The parameters of the source distributions are estimated with an Expectation-Maximization algorithm which is run jointly with the syndrome-based turbo and LDPC decoders. The case of Finite Memory Sources (FMS) has then been addressed where the sequence of symbols is assumed to be generated by an ergodic probabilistic finite state process. The distribution of the source is assumed to be dependent only on the current state of the model. A two-state Gilbert-Elliott finite state process has been considered. The Gilbert-Elliott model parameters are estimated with an Expectation-Maximization (EM) algorithm. The sequence of states of the FMS and the model parameters are iteratively estimated by the decoder together with the compressed sequence of input symbols of the source X. Sources with memory represent real sources more accurately than non-uniform sources, and their application covers wide range of systems, including biometrics or sensor networks. The rate gains obtained by accounting for the nonuniformity and the memory of the source have first been assessed experimentally with theoretical sources. The probability estimators and the modified LDPC decoders have then been integrated in a transform-domain distributed video codec, showing improved PSNR versus rate performance.