Attention-Driven Underwater Image Enhancement Framework via Edge-Aware Feature Refinement

Abstract

Underwater images suffer from unique challenges during restoration, due to absorption and scattering of light. It results in significant loss of fine details, color cast, blurriness, and irregular haze. Existing methods struggle to capture the fine details in complex underwater scenarios. To tackle these difficulties, we propose a novel Edge-guided Feature Refinement (EFR) module using channel attention and integrate this in a vision transformer based architecture. It is precisely designed to recover the blurred details. We also used a window based multi head self attention driven encoder-decoder architecture to enhance both local and global details in underwater images and also reduce the computation complexity. For validation of our approach, we performed evaluations on various paired and unpaired datasets using popular underwater image restoration metrics. By introducing the EFR module, our methodology successfully achieves the performance of the state-of-the-art techniques in various metrics.