Synthesis of super-resolution images by GAN using Channel Attention

With the development of deep learning, many studies have been conducted on techniques to generate high-resolution images from low-resolution images containing different types of degradation and noise. In this paper, we proposed two super-resolution image synthesis approaches based on a mainstream algorithm ESRGAN. The first approach is to apply channel attention to the generator of ESRGAN. The second one is to add an image quality evaluation metrics Learned Perceptual Image Patch Similarity (LPIPS) to the discriminator. Channel attention has been used to obtain positive results in image classification. We expect that the combination of channel attention and convolutional neural networks could be used to generate high-quality super-resolution images. Most of the recent studies use the peak-signal-to-noise ratio (PSNR), the structural similarity index measure (SSIM), etc. as image evaluation metrics and aim to improve their scores. However, these metrics evaluate images on a pixel-by-pixel basis, which raises a gap with human evaluation results. In our experiment, we evaluated our proposed methods on two image quality indices, Naturalness Image Quality Evaluator (NIQE) and LPIPS, using a benchmark dataset. The evaluation by NIQE and LPIPS showed that our proposed channel attention-based models significantly improved the naturalness and perceptual evaluation value of images compared to previous studies. LPIPS and NIQE scores of the synthesized images were closer to the values of the original high-resolution images generated by the model proposed in this study.