IRE: Improved Image Super-Resolution Based on Real-ESRGAN

Image super-resolution (SR) is a research field focusing on image degradation techniques. The High-order Deterioration Model (HDM) implemented in Real-ESRGAN has proven more effective in simulating the degradation of real-world images compared to conventional bicubic kernel interpolation. However, images reconstructed by Real-ESRGAN suffer from two significant weaknesses. Firstly, the rebuilt image is overly smooth and suffers from substantial texture information loss, resulting in a worse performance than classical models such as SRGAN and ESRGAN. Secondly, the reconstructed images exhibit better visualization effects but are entirely different from the original image, violating the principle of image reconstruction. To address these issues, this paper presents an improved image SR model based on the HDM implemented in Real-ESRGAN. The first-order degradation modeling of HDM was removed, and only the second-order degradation modeling was kept to reduce the degree of visual deterioration. PatchGAN was used as the fundamental structure of the discriminator, and a channel attention mechanism was added to the generator’s dense block to enhance texture details in the reconstructed images. The L1 loss function was also replaced with the SmoothL1 loss function to improve convergence speed and model performance. The proposed model, IRE, was evaluated on various benchmark datasets and compared to Real-ESRGAN. The results show that the proposed model outperforms Real-ESRGAN regarding visual quality and measures such as RankIQA and NIQE. The study also indicates that PatchGAN, as the discriminator, reduces the average training time by approximately 28%.