Underwater image restoration is fundamental for enhancing the visual perception of Autonomous Underwater Vehicles (AUVs) and Remotely Operated Vehicles (ROVs). However, existing methods face a persistent trade-off: GAN-based approaches (e.g., FUnIE-GAN) can introduce physically inconsistent artifacts due to the absence of explicit optical constraints; Transformer architectures achieve global modeling but suffer from quadratic computational complexity O(N2), limiting their suitability for real-time multi-scale and 4K underwater imagery; and several recent Mamba-based methods still use physical or degradation priors mainly as auxiliary guidance. In this paper, we propose Phys-Mamba, a physics-informed selective state-space fusion network for high-fidelity multi-scale underwater image restoration. Our method introduces two core innovations: (1) we encode UIFM-inspired latent degradation variables and use them to modulate Mamba's selective scan mechanism (Δ, B, C) via a Physics Embedding Module, encouraging consistency with the UIFM formulation and helping suppress color distortion and over-enhancement; (2) we design a Dynamic Cross-Scale Fusion Module (CSFM) that adaptively aggregates multi-scale features without channel explosion, enabling detail recovery from low-resolution semantics to 4K textures while maintaining linear O(N) complexity. Extensive experiments on EUVP, UCHN, Synthetic Deep-Sea, and LSUI benchmarks show that Phys-Mamba achieves competitive restoration performance, with PSNR of 28.12 dB and SSIM of 0.915 on EUVP, while maintaining a compact model size of 8.6M parameters. On an RTX 4090 with FP16 precision and batch size 1, the model reaches 34 FPS for 4K inference. When used as a preprocessing module, it improves YOLOv8 mAP@0.5 by 20.7 percentage points in our evaluation setting. These results suggest that UIFM-inspired selective-scan modulation is a promising direction for balancing restoration quality, global modeling, and high-resolution inference efficiency in underwater robotic vision.
🌊
