A New Trellis Vector Residual Quantizer : Applications to Image

We present a new Trellis Coded Vector Residual Quan-tizer (TCVRQ) that combines trellis coding and vector residual quantization. We propose new methods for computing quantization levels and experimentally analyze the performances of our TCVRQ in the case of still image coding. Experimental comparisons show that our quantizer performs better than the standard Tree and Exhaustive Search Quantizers based on the Generalized Lloyd Algorithm (GLA). Quantization is the process of approximating a continuous amplitude signal by a digital (discrete-amplitude) signal, minimizing a distortion measure (or error). Unfortunately , xed the coding rate R, an optimal VQ requires computational and storage resources that grow exponentially with the vector dimension. Moreover, Lin in 1] showed that the design of an optimal Vector Quantizer is an NP-complete problem. The design of sub-optimal vector quantizers is an interesting alternative to scalar quantizers for applications that require good quality performances when a limited amount of computing resources is available. In this paper we present a new VQ architecture that operates recursively on the quantization residu-als. We will demonstrate the performances of this new VQ based on a trellis structure and compare it experimentally with the standard Exhaustive Search and Tree Vector Quantizers. Quantizers based on Trellis Coding were rst proposed by Fisher et al. in 2], and made use of the set partitioning ideas of Ungerboeck 3]. Fisher et al. in 2] presented results for Trellis Coded Vector Quanti-zation (TCVQ) in up to four dimensions. Wang and Moayeri in 4] used the LBG algorithm for codebook design and report results for vector dimension up to 6. Laroia and Farvadin in 5] combined scalar-vector quantization (SVQ) with trellis coding. Belzer and Villasenor in 6] presented design techniques for vector quantizers with highly symmetric codebooks that facilitate low complexity quantization as well as partitioning into equiprobable sets for trellis coding. Our quantizer is diierent from other structures proposed in literature: all the stages of our TCVRQ are used to encode the whole vector and our trellis works by removing the statistical dependence among vector components and not among distinct vectors.