Residual Networks with Channel Attention for Single Image Super-Resolution
摘要：With the development of convolution neural network（CNN）,CNN-based methods also achieve a great success for image superresolution tasks. Further, ResNet makes it possible that the network of super-resolution can be trained deeply. However, simply increasing the depth or width of the network has a scant improvement on the reconstruction quality. Therefore, we need to exploit some new mechanisms to boost the quality of the reconstructed SR images.In this paper, we utilize channel attention mechanism to rescale channel-wise features and extract the desired high-frequency information. Additionally, we add feature fusion structure into the network in order to make full use of all the extracted middle highfrequency information instead of that extracted only by the final layer. Experiments we have conducted show that the network we proposed could reconstruct high quality images with only a few parameters.