Image compression by texture modeling in the wavelet domain

High-quality image compression algorithms are capable of achieving transmission or storage rates of 0.3 to 0.5 b/pixel with low degradation in image quality. In order to obtain even lower bit rates, we relax the usual RMS error definition of image quality and allow certain "less critical" portions of the image to be transmitted as texture models. These regions are then reconstructed at the receiver with statistical fidelity in the mid- to high-range spatial frequencies and absolute fidelity in the lowpass frequency range. This hybrid spectral texture modeling technique takes place in the discrete wavelet transform domain. In this way, we obtain natural spectral texture models and avoid the boundary blending problems usually associated with polygonal modeling. This paper describes the complete hybrid compression system with emphasis on the texture modeling issues.