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

Image degradation technique has been the focus of current research in image super-resolution( SR). The High-order Deterioration Model(HDM) of Real-ESRGAN is more effective than the conventional bicubic kernel interpolation in simulating the degradation of real-world images. However, the images reconstructed by Real-ESRGAN suffer from two notable weaknesses. Firstly, the rebuilt image is overly smooth and suffers from substantial texture information loss, making it worse than classical models like SRGAN and ESRGAN. Moreover, despite the reconstructed images having a better visualization effect, they are entirely different from the original image, violating the principle of image reconstruction. In order to address these two issues, an improved model built on Real-ESRGAN (IRE) is presented in this work. We remove the first-order degradation modeling of HDM and keep its second-order degradation modeling to lessen the degree of visual deterioration. Then we utilized PatchGAN as the fundamental structure of the discriminator, added a channel attention mechanism to the dense block of the generator, and increased the texture detail in the reconstructed images. Finally, we replaced the L1 loss function with the SmoothL1 loss function to improve the convergence speed with better model performance. The ablation study shows that the average training time was decreased by about 28% when employing PatchGAN rather than U-net as the discriminator. In addition, the ablation study presents that the RankIQA measures assessed by IRE on Set5, Set14, and DIV2K100 were decreased by 15.20%, 20.46%, and 10.27%, respectively, and the PI measures considered by IRE+ on BSD100, Urban100 were decreased by 15.76%, 8.67%, respectively, compared to Real-ESRGAN. Also, the comparison experiment illustrates that the NIQE measures assessed by our suggested model on Set5, RealSR-Canon, and RealSR-Nikon were decreased by 12.31%, 3%, and 2.41%, respectively, compared to Real-ESRGAN.