Hybrid multi-resolution analysis and weighted averaging of overcomplete orthogonal transform scheme for digital image denoising

Noise reduction is a crucial research topic of digital image quality enhancement in both theoretical and applied perspectives, and attracts extensive research efforts in decades. Weighted averaging with overcomplete orthogonal transform (WAOOT) has shown its ability to effectively remove i.i.d. noise, while maintaining edge sharpness. In this paper, the weights of the overcomplete transform set are showed to be dependent on the noise covariance matrix for non-i.i.d. noise in real digital images, a Gaussian covariance model is proposed to describe the noise correlations, and accordingly a Gaussian pyramidal multi-resolution analysis architecture is built to decorrelate the non-i.i.d. noise and reduce it by utilizing WAOOT algorithm at each layer. The simplified solution of WAOOT algorithm is further refined by global hard threshold adaption iterations and transform block size discrimination on edge pixels. Simulation results show that the proposed scheme achieves substantial improvements in both objective and subjective denoised image quality over state-of-the-art algorithms.