作者
Mochamad Rizal Fauzan,Irgi Surya,Jaja Kustija,Resa Pramudita,Rufaidah Kamilia Ahsani
摘要
Underwater images are degraded by wavelength-dependent attenuation, backscatter, low contrast, and color distortion, which can reduce visual quality and downstream reliability. This study presents PhysGuide-Net, a physics-guided restoration framework that couples degradation-aware latent estimation, physics-conditioned reconstruction, and task-feature preservation. The Degradation-Aware Physics-Guided Estimation Module predicts transmission, RGB backscatter, and attenuation-related cues and produces an explicit physics-recovered prior for a 13-channel U-Net backbone. After a source-code audit, all variants were retrained with corrected gradient flow, bounded SSIM, controlled initialization, and checkpoint-based baseline evaluation. On UIEB, PhysGuide-Net achieves 22.97 ± 3.45 dB PSNR and 0.904 ± 0.062 SSIM, significantly outperforming the unprocessed input, CLAHE, histogram equalization, and UWCNN in paired tests after Holm correction, as the difference from a controlled plain U-Net is not significant. Without retraining, it obtains 21.20 ± 4.19 dB PSNR and 0.850 ± 0.109 SSIM on LSUI. The inference network contains 9.821 million trainable parameters, requires 32.164 GFLOPs, and runs in 13.54 ms at 256 × 256 resolution on an RTX 5060. These results support the feasibility of the physics-conditioned design. At the same time, the ablation study shows that the task and physics losses do not uniformly improve reference metrics under the current weighting and that direct benchmarking against additional recent state-of-the-art models remains necessary.