Abstract: Aiming at the problem that the current convolutional neural network could not make full use of the shallow feature information, 
and it was difficult to capture the dependency between each feature channel, resulting in the loss of high-frequency information, we proposed a new generative adversarial network for image super-resolution reconstruction. Firstly, the WDSR-B residual block was introduced into the generator to fully extract the shallow feature information. Secondly, the GCNet module and pixel attention mechanism were combined into the generator and discriminator to learn the importance and high-frequency information of each feature channel. Finally, using spectral normalization instead of batch normalization that was not conducive to image super-resolution, and reduced computational overhead and stabilize training. Experimental results show that compared with other classical algorithms, the proposed algorithm can effectively improve the utilization of shallow feature information, better reconstruct the detailed information and geometric features of the image, and improve the quality of super-resolution images.

Key words:  , image super-resolution, generative adversarial network, attention mechanism, residual network