Recovering JPEG Compression Loss via Deep Learning-Based Super Resolution Techniques

Abstract

This study investigates the use of JPEG, a lossy compression algorithm, for efficient utilization of channel bandwidth. To reduce losses caused by the removal of high-frequency components in compressed images, super-resolution models are developed using deep learning methods. Specifically, the SRCNN, VDSR, and SRDenseNet models are trained from scratch using compressed images. The effects of different compression ratios on images are analyzed in terms of their impact on the recovery of high-frequency components by the super-resolution models. The performance of the models is evaluated using PSNR and SSIM metrics, considering various compression ratios.