Joint denoising and decompression using CNN regularization

Mario Gonzalez, Javier Preciozzi, Pablo Musé, Andrés Almansa

Abstract: Wavelet compression schemes such as JPEG2000 may lead to very specific visual artifacts due to quantization of noisy wavelet coefficients. These artifacts have highly spatially-correlated structure, making it difficult to be re- moved with standard denoising algorithms. In this work, we propose a joint denoising and decompression method that combines a data-fitting term, which takes into account the quantization process, and an implicit prior learnt using a state-of-the-art denoising CNN.