Jointly optimal classification and uniform threshold quantization in entropy constrained subband image coding
Are Hjørungnes, John M. Lervik · 2002
A method for coding a source modeled by an infinite Gaussian mixture distribution is proposed. The source is first split into N classes. The samples of each class are then quantized by an infinite-level uniform threshold quantizer followed by an entropy coder designed for each class. The problem of joint optimization of this system's rate distortion performance is first solved theoretically, assuming an exponential mixing density. A comparison to a system optimal for high rates, using one common quantizer for all classes, showed that for a fixed distortion the rate was reduced by 11-12% at low rates for a fixed N=5. A subband image coder, using the optimum theoretical parameter values was simulated. The resulting coder has high performance and low complexity.