Locally optimal joint encoding of image transform coefficients

We address the choice of encoder for conditional entropy-constrained trellis-coded quantization (CECTCQ), applied to image transform coefficients. The optimal CECTCQ encoder requires an (utterly intractable) exhaustive search and the standard method of greedy, sequential encoding of the coefficient "sources" is suboptimal. Alternatively, we suggest a locally optimal encoding algorithm, guaranteed to improve performance over greedy encoding, and yet with manageable increases in encoding complexity. This method uses dynamic programming as a local optimization encoding "step", repeatedly applied until convergence. Simulations demonstrate up to 1.5 dB gain over greedy CECTQ encoding of block-transformed images.