DC-Deblur: A Dilated Convolutional Network for Single Image Deblurring

Single image deblurring is a significant and challenging task in image processing vision and machine learning. Convolutional Neural Network (CNN) based models for deblurring often have a complex structure and a considerable number of parameters compared with those for other image restoration tasks such as image denoising, dehazing and super-resolution. The main reason is the requirement of large reception fields, which are important to image deblurring due to possible large blur kernels. Dilated convolution is a useful way to increase reception field without adding extra parameters. In this paper, we propose a novel network by adopting a dilated convolution structure, and we further improve the training process by combining L1 loss, MS-SSIM loss and MSE loss. The proposed network is light and fast. Quantitative and qualitative experiments indicate that our method outperforms state-of-the-art models, in terms of performance and speed.