LBCRN: lightweight bidirectional correction residual network for image super-resolution

Shuying Huang, Jichao Wang, Yong Yang, Weiguo Wan, Guoqiang Li

Abstract: Currently, single image super-resolution methods based on convolutional neural networks have achieved remarkable results. These methods typically increase the depth and complexity of a network to improve its performance, which increases the network’s computational burden. To solve these problems, this paper proposes a new lightweight bidirectional correction residual network (LBCRN) for image super-resolution. LBCRN includes dominant correction and return correction modules. In the dominant correction module, a feature fusion residual block (FFRB) with fewer parameters is constructed to learn and fuse the extracted features of different layers. Meanwhile, to restore the important features while keeping the network lightweight, a hybrid-attention block (HAB) is also proposed based on the channel and spatial attention mechanisms. Finally, to better constrain the network training, a return correction module is designed by simulating the process of image degradation. Experimental results on multiple datasets show that the proposed LBCRN has better subjective performance and quantitative results than most existing lightweight networks.