A Wavelet-Domain Hidden Markov Tree Model with Localized Parameters for Image Denoising

Abstract

Wavelet-domain hidden Markov tree models have been popularly used in many fields. The hidden Markov Tree (HMT) model provides a natural framework for exploiting the statistics of the wavelet coefficients. However, the training process of the model parameters is computationally expensive. In this paper, we propose a HMT model with localized parameters which has a fast parameter estimation algorithm with no complex training process. Firstly, Wold decomposition is used to reduce the influence on the estimation of image noise variance due to texture. Secondly, coefficients in each subband are classified into two classes based on spatially adaptive thresholds. Thirdly, parameters of different class are estimated using the local statistics. Finally, the posterior state probability is estimated with an up-down step like the traditional HMT model. We apply this model to image denoising and compare it with other models for several test images to demonstrate its competitive performance.