Enhancement of underwater images using GAN and Refinement network

摘要

A refinement network is a need to further improve the generated images by fine-tuning the output, focusing on enhancing finer details and textures, especially in challenging underwater environments. The enhancement of aquatic visuals is critical challenge in computer vision due to the inherent difficulties posed by low-light conditions and the scattering of light underwater. This paper proposes a novel approach to enhance the quality deep sea images using GAN integrated alongside a refinement network. The proposed method leverages a three-stage architecture consisting of a Generator, Discriminator, and a Refinement Network to deal with the concern of dreadful visibility, color distortion, and low contrast in deep-sea images. The generator is designed to learn the mapping between the degraded input images and their enhanced counterparts through a series of layers, whereas the Discriminator works to differentiate between real and fake (generated) images to guide the generator’s learning process. The model is trained using a custom dataset consisting of paired underwater images and their enhanced versions, with performance evaluated using metrics. The results highlight that the integration of the refinement network considerably improves the visual quality and fidelity of the generated images, outperforming traditional image enhancement methods. Additionally, the model shows robustness in handling varying underwater conditions, making it applicable for real-world applications in marine biology, underwater robotics, and environmental monitoring.