Practical rate-control tools for wavelet-based image compression

Transform-based compression algorithms have long been used to represent images efficiently. Recent advances in wavelet quantization strategies and coding methodologies have resulted in higher quality images compressed under a rate constraint. For real-time compression, or implementation on devices with limited resources, ideal rate-control algorithms are simple, accurate and fast; for flexibility, they must be robust to choices of quantization scheme and coding mechanism. Several rate-control algorithms are presented, varying in performance and complexity, in a wavelet-based compression framework.; The key to each approach is the ability to accurately estimate the coded rate associated with a collection of quantized wavelet transform coefficients. A fast estimation technique is derived for this purpose, based on a versatile doubly stochastic probability model. The amount of overhead computation required to fit this model is incrementally greater than that required to compute raw subband moments, and coded rate is estimated simply by evaluating a set of low-degree polynomials. A modified version of Newton's method is combined with this estimation scheme to efficiently determine coder-specific behavior in a robust two-pass rate-control algorithm. Both this algorithm and a novel heuristic alternative outperform other state-of-the-art approaches in terms of speed, accuracy, robustness and flexibility. A scalable rate-control algorithm that generates a layered representation of transform data is also discussed. The order of the layers is determined by using an arbitrary quantization strategy to control step-sizes induced in the reconstructed image; this order can be optimized for performance over a non-contiguous collection of discrete rates or a continuous range of rates.; Results are extended to more sophisticated spatially-adaptive compression schemes requiring explicit side information. These include a wavelet-based approach that quantizes each individual coefficient with a different step-size using psycho-visual masking principles, and a distributed spatial domain algorithm that codes multiple degraded versions of the same scene, for enhancement post-processing. New efficient coding strategies are introduced that successfully mitigate the side information cost. The rate-control tools developed in this thesis yield highly accurate performance when used with non-spatially-adaptive state-of-the-art compression schemes; similar accuracy results with the proposed spatially-adaptive schemes, despite the inclusion of new dependencies in the compression framework.