Evaluation of REAL-ESRGAN Using Different Types of Image Degradation

This study evaluates the performance of the REAL-ESRGAN [1] model on images with varying levels of degradation using the DIV2K dataset [2], such as the Wild, the Mild, the Difficult, and the x8 subsets. REAL-ESRGAN was created to solve super-resolution problems and aims to produce high-resolution images from low-resolution images. Experiments were conducted at scales of x2 and x4, and performance was measured using Full-Reference metrics (LPIPS, PSNR, SSIM) and No-Reference metrics (NIQE, MANIQA, CLIPIQA, and PI). The Results were good, especially with the x2 scale; it has higher PSNR and SSIM scores, lower LPIPS and NIQE values, and enhanced visual and perceptual quality. The model faced more significant challenges with the wild and the difficult datasets because they have more complex degradations and compression artifacts; it can be seen with unstable results of Full-Reference and No-Reference metrics. On the contrary, the Mild and x8 datasets yielded better results in both metrics; not only that, even the computational cost for Mild and x8 outperforms the rest of the dataset. This study shows the strengths and limitations of REAL-ESRGAN in handling different levels of image degradation. For future research, the model needs enhancement to tackle the degradation format of the wild and the difficult dataset. It would be good if the REAL-ESRGAN improvement could also maintain the computational cost.