EGNet: enhanced gradient network for image deblurring

In recent years, convolutional neural networks have been increasingly developed in the field of image deblurring. However, there are still many problems in the existing methods. First, most methods lack attention to the middle layers of the networks, resulting in the lack of gradient information which is important but cannot be restored well by the existing methods. And the ability of the networks which can reconstruct sharp images is reduced. Second, the existing deblurring methods cannot perform the self-ensemble of the networks, which has been proved to further improve the network performance in many fields such as image classification and single image super-resolution. Third, the loss function of the existing networks all directly use MAE loss, MSE loss or adversarial loss, which lack the special design for edge information and structural information in the images. To solve the above problems, we enhanced the gradient information of the feature maps after the encoder in the network and then designed the self-ensemble network to further improve the performance; finally, we redesigned the loss function to adaptively redistribute the weight of each position in the gradient map according to the input image. Through the extensive experiments, our proposed method achieves the state-of-the-art results of image deblurring on the existing dataset and also outperforms other methods in generalization ability on RealBlur dataset as well as real blurry images.