Rapid Image Super Resolution

Abstract

Image Super-Resolution (ISR) is a long-established challenge that finds extensive usage in the field of medical imaging, media consumption, drone surveillance, etc. Recent advancements in deep learning and improved GPU hardware have enabled researchers to create sophisticated research work. Earlier approaches focused on improving the Peak-Signal-to-Noise-Ratio of SR images, but it led to the loss of finer details. Recently GAN-based architectures like SRGAN and ESRGAN have been introduced which improves the human-perceived quality of the generated SR images. However, these architectures have high computational costs, not suitable for low-end or mobile devices. We propose a lighter, faster, and optimized GAN-based super-resolution architecture, Rapid-SR, using depthwise convolutional layers. It produces similar results as state-of-the-art approaches while reducing the model parameters and the time taken to produce SR images substantially. We also use a novel training strategy for Rapid-SR which incorporates the measure of the perceived similarity in the training loss by using Learned perceptual image patch similarity (LPIPS). The results are analyzed and compared using PSNR/SSIM, LPIPS, and Mean Opinion Scoring.