Underwater Image Enhancement Using a U-Net Enhanced GAN with Multi-Scale Feature Fusion and Channel-Spatial Attention

Abstract

This study proposes an innovative approach to underwater image processing to counteract the prevalentissue of detail loss associated with current techniques. By enhancing the discriminator within thegenerative adversarial network through the incorporation of a U-Net, the method aims to preserve theintegrity of generated image details. Subsequently, it integrates multi-scale feature fusion and spatialchannel attention mechanisms into the collaborative design of the image processing framework toreinforce the expression and reconstruction capabilities of key features. Ultimately, performancevalidation and analysis were carried out on the Underwater Dark, Underwater ImageNet, andUnderwater Scenes datasets, with the primary evaluation metrics including PSNR, SSIM, and Fréchetdistance. The experimental results demonstrated that the SSIM of the proposed method across the threedatasets was 0.82, 0.85, and 0.83, respectively, which represents an improvement of 0.17, 0.12, and 0.11over the unimproved methods. Its PSNR maintained the highest values of 28.9, 30.1, and 29.8 across allthree datasets, showing an enhancement of 7.9, 7.6, and 7.8 over the baseline models. The Fréchetdistance remained the lowest at 22.0, 19.0, and 20.0, indicating a reduction of 21.0, 18.0, and 19.0compared to the baseline methods. These results indicate that the proposed image processing techniquebased on U-Net and multi-scale feature fusion effectively solves the problem of detail loss in underwaterscenes, has strong generalization ability, and provides feasibility for complex underwater imaging tasks.