Deep Underwater Image Enhancement using RGB ResUNet and Hybrid Perceptual-Structural Loss

Underwater images typically suffer from issues like scattering, absorption, and notable color distortions. These issues reduce their clarity and value in marine research. In this study, we used a deep learning model to enhance these images. The model is trained on the UIEB paired dataset with an encoder-decoder network in the RGB domain. We use a hybrid loss function to guide the learning process. This function is a combination of Mean Squared Error, Structural Similarity Index, and perceptual loss. This helps the model maintain pixel level accuracy, keeps structural details, and improves visual quality. Our experiments show that this method successfully enhances underwater images. It restores natural colors, contrast, and finer details. Currently, the model focuses on RGB inputs only. In future work, we plan to include dual color-space processing with HSV features, attention-guided fusion, and applications like coral reef detection and real-time deployment.