Abstract:Despite the rapid development of single-image super-resolution (SISR) methods of generative adversarial networks (GAN), which can reconstruct visually realistic images, the problem of high discrepancy between the recovered details or textures and the ground truth persists. To address this issue, an SISR reconstruction GAN based on a feedback and attention mechanism (FBSRGAN) is proposed. Specifically, we select a network with a feedback mechanism as the generator, which can gradually create high-resolution images through the feedback connection. The attention mechanism is combined with the feedback block to adaptively select useful feature information and effectively process the feedback stream and enhance the image output quality. We use the relative average least squares GAN loss to reduce the instability of the optimization generator process to guide GAN to obtain more realistic results. The results show that compared with the ESRGAN method, when the amplification factor is 4, PSNR and SSIM of the proposed method increase by 0.386 and 0.0141, respectively, and PI decreases by 0.284, while the number of parameters is only 18.5% of that of ESRGAN, when tested on the Set5 dataset. Compared with existing GAN-based SR methods, FBSRGAN achieves superior performance in terms of both perceptual ability and image distortion.

Image super-resolution reconstruction based on generative adversarial network model with feedback and attention mechanisms