Single Image Super-Resolution Based on Self-Attention

Single image super-resolution (SISR) is a challenging task and has collected extensive attention in both industry and academia. The most challenging problem in super-resolution is how to recover the finer texture details. And we find the generated images usually have low-scale contrast. In this paper, we present a novel super-resolution method based on generative adversarial networks (GAN). Our model is based on SRGAN, instead, we remove most of the batch normalization layers in generator to get higher-scale contrast images and accelerate training speed. Because batch normalization layers can get rid of range flexibility from networks, this causes the generated images have lower-scale contrast compared to the origin images. We also add the self-attention module into generator to get more global dependencies when convolution operations can capture enough local dependencies but little global dependencies. We take advantage of both local dependencies and global dependencies to improve super-resolved texture details and structural, we call our model SASRGAN. The images generated by our model have higher Structural Similarity Index Measure (SSIM), it proves that our model has available performance.