Underwater image geometric distortion and detail degradation caused by water absorption and scattering, uneven illumination, and refractive distortion severely limit the accuracy of underwater vision tasks. To this end, this paper proposes an E-PUGAN enhancement model based on a physics-model-guided GAN framework. Its core innovations include: (1) designing a position-aware EfficientViM module based on the HSM-SSD architecture, which enhances spatial structural understanding by embedding spatial coordinate information and solves the problem of local structural misalignment caused by underwater refractive distortion; (2) combining a channel/spatial dual-gating mechanism to dynamically adjust the feature weights of degraded regions, and jointly utilizing the dual-path fusion module DUF to achieve adaptive multi-granularity feature fusion, addressing the over- and under-enhancement issues in traditional residual network connections; (3) A triplet dataset for underwater images, UIE-Triplet160, is proposed, and the UIE-LPIPS loss function is constructed to improve the alignment between underwater color distortion correction and human visual perception. Experiments on four mainstream benchmark datasets, including UIEB and LSUI, demonstrate that E-PUGAN shows significant advantages over eight state-of-the-art methods, achieving objective metric improvements of up to PSNR 28.02 dB and SSIM 0.906. Visualization results further verify the superiority of the proposed method in detail preservation and natural color restoration.
Song et al. (Tue,) studied this question.