A Quantization Scheme for Modeling and Coding of Noisy Texture in Natural Images

Noisy textures in images pose a particular problem to image compression as they are not well suited for decorrelation. Established image compression methods thus need to be driven at relatively high bit rates in order to achieve visually transparent compression of this type of signal. At the same time, it is an intuitive observation that pixel-wise exactness of the reconstruction of noisy texture is not necessary in order for it to be transparent. Thus, it is desirable to find a more efficient representation of such texture. In this paper, we suggest the use of a denoising algorithm to achieve a conservative decomposition of the image signal into an “exact” structure component and a “statistical” noise component. We develop a noise model which allows us to represent and parameterize noisy textures as a nonstationary ARX process. Furthermore, we present an estimation algorithm for the model parameters and evaluate results on a number of well-known images.