Simplification of VLC tables with application to ML and MAP decoding algorithms

Many source coding standards (JPEG, H263+, H264), rely heavily on entropy coding with variable-length codes (VLC). However, bitstreams made of VLC-encoded data are particularly sensitive to transmission errors. Recent results tend to use knowledge of the VLC structure in order to perform an efficient decoding of the bitstream. These techniques use a trellis describing the structure of the VLC codebook and assume that some a priori information is available at decoder side. Significant improvements, compared with prefix decoding of bitstreams are achieved. However, the complexity of these techniques may become intractable when realistic VLC codebooks are considered. This paper presents an algorithm for compacting VLC tables. The codewords are grouped into a minimum number of classes. Decoding algorithms may then work on a reduced number of classes, instead of working on the whole set of codewords. A proof of optimality is provided for the VLC table-compaction algorithm. The algorithm is applied to the H263+ VLC codebook and merges the 204 codewords into 25 classes. The resulting compact tables are shown to be exactly equivalent to the initial ones when used with hard decoding algorithms. The properties of the associated soft decoding algorithms using these compact tables are also evaluated.