A Dual Residual Network with Channel Attention for Image Restoration

Abstract

Deep learning models have achieved significant performance on image restoration task. However, restoring the images with complicated and combined degradation types still remains a challenge. For this purpose, we proposed a dual residual network with channel attention (DRANet) to address complicated degradation in the real world. We further exploit the potential of encoder-decoder structure. To fuse feature more efficiently, we adopt the channel attention module with skip-connections. To better process low- and high-level information, we introduce the dual residual connection into the network architecture. And we explore the effect of multi-level connection to image restoration. Experimental results demonstrate the superiority of our proposed approach over state-of-the-art methods on the UDC T-OLED dataset.