Wavelet-based Hidden Markov Tree (HMT) models have been proven to be useful tools for statistical signal and im- age processing. But wavelet transform lacks directional in- formation, so it is difficult to get better performance in many image processing tasks with complicated texture images. In this paper, we use the nuqDFB to decompose an image into twelve highpass subbands and one lowpass subband, while remaining perfect reconstruction and maximally decimation property. By repeating the same decomposition to the low- pass subband, a quardtree structure of the coefficients is established that can be used in the training of HMT mod- els. Then, we adopt a new method via contexts to exploit the intra-scale clustering property and implement it in the HMT training iteration process. To illustrate the power of this approach, this algorithm is used in image denoising. The results show that it is obviously superior in both vision and Pulse Signal to Noise Ratio (PSNR).