Exploring Loss Functions, Metrics, and Datasets for Super Resolution Deep Learning Models

Single Image Super Resolution (SISR) is an active area of research within the deep learning community, aiming at enhancing the resolution of a low-resolution (LR) observation to produce its high-resolution (HR) counterpart. We begin by examining various loss functions employed in SISR models, analyzing how they affect performance metrics commonly used to assess these methods. The evaluation of different metrics on the quality and accuracy of the generated HR images is critically evaluated using HR generated samples. Then we explore the significance of the training datasets, namely how the specificity of these datasets impacts the performance of SISR models. Both quantitative metrics and qualitative assessments of the HR samples are used to gauge the effectiveness of different training datasets. Finally, the paper compares two distinct SISR approaches: the supervised approach, where models are pre-trained on extensive datasets before inference, and the unsupervised approach, which generates HR images from a single LR image without prior training.