Zhang Kai Zuo Wangmeng Chen Yunjin Meng Deyu Zhang Lei

Abstract

Discriminative model learning for image denoising has been recentlyattracting considerable attentions due to its favorable denoising performance.In this paper, we take one step forward by investigating the construction offeed-forward denoising convolutional neural networks (DnCNNs) to embrace theprogress in very deep architecture, learning algorithm, and regularizationmethod into image denoising. Specifically, residual learning and batchnormalization are utilized to speed up the training process as well as boostthe denoising performance. Different from the existing discriminative denoisingmodels which usually train a specific model for additive white Gaussian noise(AWGN) at a certain noise level, our DnCNN model is able to handle Gaussiandenoising with unknown noise level (i.e., blind Gaussian denoising). With theresidual learning strategy, DnCNN implicitly removes the latent clean image inthe hidden layers. This property motivates us to train a single DnCNN model totackle with several general image denoising tasks such as Gaussian denoising,single image super-resolution and JPEG image deblocking. Our extensiveexperiments demonstrate that our DnCNN model can not only exhibit higheffectiveness in several general image denoising tasks, but also be efficientlyimplemented by benefiting from GPU computing.