Comparative analysis of modern ESRGAN models in the ill-posed problems of blind super resolution and motion-blurred /low-light image restoration

Abstract

This work explores the task of super-resolution using advanced GAN-based architectures tailored specifically for this purpose. A primary focus is addressing the inherent challenges of the ill-posed nature of super-resolution by incorporating synthetic images in various ways, emphasizing the importance of model selection that integrates synthetic image approaches. Additionally, we investigate the relationship between super-resolution and related image reconstruction tasks such as image deblurring and low-light image enhancement through transfer learning. Our experiments reveal the capabilities of state-of-the-art super-resolution models while also highlighting their potential and limitations, particularly in handling cues associated with deblurring and low-light scenarios. Despite their high performance in super-resolution, these models struggle with the complexities of upscaling processes inherent in these tasks, underscoring the necessity for pretraining on specialized datasets to address these challenges effectively. This study contributes insights into the strengths and limitations of advanced super-resolution techniques and emphasizes the importance of tailored training for addressing specific image enhancement tasks.