Underwater Image Super-Resolution Based on the Combination of Generative Adversarial Networks and Transformer

Advancements in technology have made it easy to put equipment and gear into the water for photography purposes. Nevertheless, there are numerous additional obstacles while capturing photographs beneath the water’s surface. These factors encompass light asymmetry and subpar image quality caused by distortion, refraction, or plankton presence. We conducted research on techniques to optimize image quality and boost color accuracy. The research is focused on applying and improving various models, as well as training and evaluating the acquired outcomes. Regarding the first goal of increasing resolution, we collect and aggregate different datasets together, training on two main frameworks: Generative Adversarial Networks (GAN) and Transformer. Specifically, the selected models include Deep Blind Image Super-Resolution (BSRGAN), Transformer for Single Image Super-Resolution (ESRT), and Image Restoration Using Swin Transformer (SwinIR). Second, with the issue of color improvement, we assume the condition that the photo has a spectral color range that does not lose too much but has unevenness between color layers, so the color distribution can be changed to be more balanced. Specifically, K-Nearest Neighbors (KNN) is the main method applied for white balance application on images. Finally, we try to combine both options to get the best results, compare the performance between the models, and give an assessment of them through International Qualifications Assessment (IQA) indicators.