Multi-level Feature Extraction and Edge Reconstruction Fused Generative Adversarial Networks for Image Super Resolution

Abstract

At present, the image super-resolution method based on convolutional neural network has achieved a very high PSNR, but the high-frequency information obtained by using the mean square error as the loss function is not sufficient, and when the scale factor is large, the detail texture of the restored image is blurred, and it is not completely consistent with the human visual perception. Therefore, this paper proposes an image super-resolution algorithm based on GAN. We modify the residual block of the original SRGAN generator network into three modules: Edge-Reconstruction network, Low-Frequency feature (LF-feature) extraction module and Residual network. The Edge-Reconstruction network reconstructs the edge of SR image, and the LF-feature extraction module extracts the low-frequency information of the image. After that, the two parts of information are fused and transmitted to Residual network to extract the high-frequency information of the image, and then the SR image is reconstructed and enlarged. And use skip connection in the network to increase the network depth. The training results show that our network has better performance in both objective evaluation indicators and subjective vision. Even with a large-scale factor, our network can recover fine texture information.