Research on Residual Learning of Deep CNN for Image Denoising

Abstract

Image denoising is a classical but still popular research topic. Removing noise from corrupted images is an indispensable step for many practical applications. Deep Learning for image denoising has shown favorable performance. Residual Learning of Deep CNN (DnCNN) is proposed for image denoising, and shows desired performance. In the paper, based on DnCNN, some hyperparameters are adjusted for better performance. In addition, a validation step is added during the training process, which allows us to observe the training process to avoid overfitting. With the validation step during the training process, a novel method of learning rate adjustment is introduced to help train the best model for the network. The results show the adjusted network has a better performance compared to the baseline of DnCNN.