A Multi-scale Dilated Residual Convolution Network for Image Denoising

Due to the excellent performance of deep learning, more and more image denoising methods based on convolutional neural networks (CNN) are proposed, including dilated convolution method and multi-scale convolution method. A fundamental issue is how to obtain multi-scale information and to recover the image detail. In order to solve the issue, we present a multi-scale dilated residual convolution network (MDRN), which has a multi-scale feature extraction block and dilated residual block. The multi-scale feature extraction block, making full of the multi-scale information, is presented by incorporating multiple-scale pixel shuffle downsampling, which can extract salient features from input images. At the same time, the dilated residual block expands the receptive field and can effectively utilize the global image information. Extensive experimental results on both the synthetic and real-world noisy images show that our method is effective and surpasses the state-of-the-art denoising methods in terms of both quantitative and qualitative evaluations.