FASRGAN: Feature Attention Super Resolution Generative Adversarial Network

The advent of multimedia-based communication requires high-speed, memory-efficient data storage of high quality images. One method for reducing the dimensions of an image while preserving the finer texture characteristics is image super resolution. This paper proposes a new model to deal with the reconstruction time and artifacts that GAN-based deep learning models face. To increase the speed, FASRGAN aims at reducing the complexity of generators. The network architecture builds on the ideas from ESRGAN and RAMS to derive a more efficient FASRGAN. This paper introduces the Residual Feature Attention Block (RRFAB) as a fundamental feature extraction block. Moreover, a relativistic discriminator is employed, inspired by relativistic GAN, that predicts the realness value of the generated image rather than a strict class value. Ultimately, training is performed on the model to improve on the content loss, which helps to converge the weights in the later training stages. The proposed model FASRGAN, achieves PSNR scores 31.5 and 31.04, SSIM scores 0.975 and 0.93 on Set5 and Set14 datasets, respectively, that are greater than the state of the art techniques and SSIM scores that are comparable to them.