A wavelet-based image compression technique using randomized quantization

This dissertation has presented a new image compression technique using a joint framework of randomized or dithered quantization and the wavelet based denoising methods. The new method is used for several images and compared with results obtained from the classical deterministic quantization model. The theory of dithered quantization is found to have many desirable properties in respect of the total quantization noise. In this work, the theory of dithered quantization is studied and extended relative to subband coding. Also the traditional bit allocation algorithm based on the minimization of L2 error norm is modified by considering the L2 error norm. A closed form equation is obtained to allocate variable bit lengths to various subbands by minimizing the L2 norm in the context of rate-distortion theory. Utilizing the properties of dithered quantization and the new bit allocation algorithm, a novel image compression technique is outlined. The new method is shown to be comparable with an image compression system using subtractive or pseudo-randomized technique but has more relevance from a practical implementation standpoint. The perceptual quality of the compressed image with the new technique is found to be better than the result obtained from the classical deterministic model employing the traditional L2 based bit allocation method.