Joint Learning of Multiple Image Restoration Tasks

Convolutional neural networks have recently been successfully applied to the problems of restoring clean images from their degraded versions. Most studies have designed and trained a dedicated network for each of many image restoration tasks, such as motion blur removal, rain-streak removal, haze removal, etc. In this paper, we show that a single network having a single input and multiple output branches can solve multiple image restoration tasks. This is made possible by improving the attention mechanism and an internal structure of the basic blocks used in the dual residual networks, which was recently proposed and shown to work well for a number of image restoration tasks by Liu et al. Experimental results show that the proposed approach achieves a new state-of-the-art performance on haze removal (both in PSNR/SSIM) and JPEG artifact removal (in SSIM). To the authors' knowledge, this is the first report of successful multi-task learning on diverse image restoration tasks.