Single Image Deblurring Based on Auxiliary Sobel Loss Function

This paper presents an effective auxiliary loss function which uses the Sobel operator to improve the performance of image deblurring methods based on deep learning. Conventional deep learning-based image deblurring methods exploit mean square error (MSE) loss function that simply measures the intensity difference in pixel-wise manner. Although recovering the lost high-frequency component is the main purpose of image deblurring, MSE loss function often fails to train the network in recovering the high-frequency components. To alleviate this issue and further improve the performance of conventional methods, we propose an auxiliary Sobel loss function which guides the network to focus on recovering the high-frequency components of resultant images. The experiment results show that the networks trained by the Sobel loss function and the conventional MSE loss function outperform the existing methods in both quantitative and qualitative evaluations.