UWEGAN: Enhancement of detailed features and restoration of image color

Abstract

Underwater images are crucial in various domains, including marine science, resource exploration, ocean engineering, and underwater surveys. However, underwater images often suffer from issues such as detail loss, color distortion, and blurring due to complex water environments. To address these problems, this paper proposes a novel underwater image enhancement algorithm named UWEGAN, which combines a U-shaped encoder with a Generative Adversarial Network. The generator in UWEGAN integrates three key modules: the Multi-scale Feature Fusion Module (MFFM), the Feature Interaction Attention (FIA) module, and the Composite Residual Extraction Unit (CREU). Specifically, MFFM is designed to extract features from different spatial levels using parallel convolutions with varying kernel sizes and then fuses multi-scale global features to enhance the network?s representation capability. To correct color distortion, the FIA module models both channel-wise and pixel-wise relationships, enabling more targeted color adjustments and improving the overall color balance of the image. For image deblurring, the CREU replaces traditional convolution blocks with densely connected residual units that utilize deep residual learning and multi-level feature extraction strategies. This allows the network to effectively differentiate between noise and real structural information, thereby preserving image details. Extensive experiments conducted on public underwater datasets confirm that the proposed method significantly improves visual quality. Quantitative evaluations show that UWEGAN achieves average improvements of 2.33%, 2.12%, and 1.67% in PSNR, SSIM, and MSE, respectively, demonstrating its effectiveness in enhancing underwater images under challenging conditions.