Fusion of Reinforcement Learning and Residual Correction for Underwater Image Enhancement

Underwater images often exhibit degradation problems, including color distortion, reduced contrast, and blurred details due to light absorption, scattering, and complex water conditions. These problems hinder the observation and analysis of underwater scenes. Conventional enhancement techniques often struggle to address the complexity of these degradations and rely on pairs of reference images. To overcome these limitations, we propose a new underwater image enhancement framework that uniquely incorporates reinforcement learning (RL) and residual correction mechanisms. By describing the enhancement process as a reinforcement learning task, our approach enables intelligent agents to dynamically adapt and optimize enhancement strategies. The intelligent agent uses perceptual, color, and residual features to guide its decision-making, and residual correction provides real-time feedback on the effectiveness of the enhancement. This integration not only enhances the adaptation of enhancement operations to different underwater environments but also improves visual quality and structural fidelity. Our approach does not require large amounts of paired training data and exhibits better performance compared to existing model-based and deep learning approaches.