Manipulating rates, complexity and error-resilience with discrete transforms

Abstract

The common conception of transform coding is a a computationally efficient alternative to vector quantization. At high rates, it is not the partitioning itself but the efficiency of the scalar entropy coding which makes transform coding useful. With this view a class of discrete transforms derived from linear transforms are used to pursue three objectives: reducing coefficient entropies (getting coding gain as in conventional transform coding), reducing the complexity of entropy coding (allowing many coefficients to be efficiently coded with identical entropy codes), and having robustness to coefficient erasures.