LAUSR

In recent years, the methods based on deep convolutional neural networks (DCNN) have greatly promoted the development of image super-resolution (SR). However, deeper and wider SR networks are more difficult to train. For the SR task, the low-frequency information contained in low-resolution images is very important, and the neglect of exploring the feature information across channels hinders the representational capability of DCNN. To address these problems, we enhance the representational capability of DCNN by utilizing the interactive relationship among multiple channels. In this paper, a residual attention network using multi-channel dense connections (MCRAN) is proposed to improve the image super-resolution significantly. This method can make full use of multi-channel information for more effective feature expression and learning. In MCRAN, a multi-channel residual attention (MCRA) module is designed to coalesce the features of multiple different channels and the attention mechanism is applied to adjust the channel features adaptively. Accordingly, the channel features own more discriminative representation. In addition, the multi-source residual group (MSRG) structure is developed to construct a deeper network and simplify the training of network, which contains several long non-local skip connections (L-NLSC) to capture global low-frequency information in remote space. Besides, MSRG contains some short local-source skip connections (S-LSSC) to enhance the information interaction of local network. Extensive experimental evaluation on benchmark datasets on single image super-resolution proves the superiority of the proposed MCRAN.