Image Denoising Via Multi-Task Learning

Although convolutional neural networks (CNN) have notably improved the effect of image denoising, removal of non-Gaussian noise remain a challenging problem. In this work, the statistical characteristic of image residuals is investigated and used as auxiliary information for better removing complex type noise via multi-task learning method. We propose an improved algorithm for denoising CNN (DCNN) by optimizing the training of the DCNN and it can achieve a Pareto optimal solution. Extensive experiments on benchmark data sets with different noise models demonstrate that the proposed method can effectively improve the quality of denoised images both in Gaussian and non-Gaussian noise, even when the network architecture is left unchanged.