A GPU parallel Algorithm for Image denoising based on Wavelet Transform coefficients Thresholding

Denoising using Discrete Wavelet Transform is very popular in image processing. The main tasks of this kind of approach include the image decomposition into random sub-bands and the Wavelet coefficients extraction. In this paper a GPU parallel algorithm to compute and extract Wavelet coefficients is proposed. As preprocessing we use the Donoho method to study the amount of noise present in the images; while for suppression or shrinkage criterion we suggest the Garrote threshold. The selection process in this work was performed through two algorithms: Hard and Non-Negative Garrote Thresholding. Our work-frame uses a parallel approach to extract Wavelet coefficients at level of sub-bands, so that significant improvements are achieved over the CPU version, which a speedup gain of around four times with respect to the serial version. Some tests and numerical experiments show the efficiency and the quality obtained using the proposed method.