Joint Optimization in Underwater Image Enhancement: A Training Framework Integrating Pixel-Level and Physical-Channel Techniques

Abstract

In recent years, with the increasing interest in marine research, the need to collect and process clear underwater optical images has become crucial. However, underwater images suffer from the absorption and scattering effects of the environment. In this paper, we propose Hybrid Underwater Image Enhancement Network (HUWIE-Net), a novel deep learning-based underwater image enhancement framework consisting of three distinct sections, which include an Image-to-Image Module, a Physics-Informed Module and a Fusion Module. The training methodology of HUWIE-Net is designed to jointly optimize both pixel-level-based and physical-channel-based enhancement modules. In this framework, while Image-to-Image Module is used for color correction in pixel level, Physics-Informed Module is used for dehazing by exploiting the underwater image formation model which defines the deformations in the underwater light propagation channel. We also propose to use the joint loss function for both Image-to-Image Module and Physics-Informed Module to enforce the joint optimization for better underwater image enhancement performance. The results of experiments conducted with real-world underwater images show that the proposed model achieves improved performance compared to state-of-the-art methods. The code for the newly developed HUWIE-Net is available at <uri>https://github.com/UIE-Lab/HUWIE-Net</uri>.