Multi-channel Network with Dense Attention for Image Super-Resolution

With the development of deep learning technology, various image super-resolution (SR) reconstruction methods based on the convolutional neural network (CNN) have been proposed. However, most of the existing methods cannot make good use of the tiny feature information in the original images and do not fuse local and global information, which results in blurred texture details of the reconstructed image. Aiming at this problem, this paper proposes a Multi-channel Network with Dense Attention (MNDA) for image super-resolution. The proposed network uses a multi-channel parallel convolution module (MPCM) to extract rich features from low-resolution images. Next, the extracted features are sent to the Long-Short path Attention Module (LSAM), which fuses and compresses the features extracted by the previous module, and the extracted feature information is distinguished by the Spatial-Channel Attention Block (SCAB), thereby enhancing the recognition ability of the network. Multiple long-short path attention modules are combined using Dense Local Connection (DLC) for image reconstruction. When performing the most difficult 4x reconstruction, on the four test datasets, the PSNR of the proposed algorithm is improved by 0.84/0.37, 0.33/0.16, 0.93/0.51, and 1.45/0.78, respectively, compared with IDN/SADN. The experimental results show that the proposed algorithm has good performance in natural image super-resolution reconstruction, and can reconstruct clearer images.