Analyzing different loss functions for Single Image Super-Resolution

Single Image Super-Resolution (SISR), which aims to recover a high-resolution (HR) image from a low-resolution (LR) one, is an ill-posed problem. Convolutional Neural Networks (CNNs) have been used in low-level vision tasks such as Super-Resolution (SR), and inspired by impressive results in high-level tasks. By choosing the proper structure, the methods can be improved significantly. In this case, selecting an appropriate loss function is essential for any deep learning task, especially in SISR. The exploited loss function impacts the quality of the images produced by the SISR algorithms. Some loss functions can make the output image look blurred or unnatural, which goes against the purpose of SR. To ensure that the output image retains the content of the original photo while also improving the structure and texture, it is essential to choose a loss that is well suited for the task. In this paper, various loss functions for SISR are reviewed. Then, we present an overall analysis of loss functions for SISR based on our exploration.