PINE: Physics-regularized neural enhancement for underwater images and frame-wise video via implicit representation

Enhancement of underwater images and video remains challenging due to spatially varying optical properties and rigid resolution constraints. We present PINE (Physics-regularized Implicit Neural Enhancement), which models restoration as a continuous estimation of a physical parameter field via coordinate-based implicit neural representations, applied frame-wise to video. PINE combines Fourier positional encodings with hierarchical image features into an implicit MLP to predict transmission and attenuation coefficients, as well as background light, at arbitrary coordinates. Rather than relying on hard physical derivations, the framework adopts differentiable soft constraints that enforce the oceanographic attenuation ordering β_r ≥ β_g ≥ β_b, enabling resolution-independent processing from training patches to native 4K without retraining. On the UVEB benchmark, PINE achieves 30.28 dB PSNR under standardized 256 × 256 evaluation and 29.28 dB at native resolution with only 1.07M parameters. Ablations confirm physics regularization is vital to prevent parameter collapse, avoiding a 5.57 dB PSNR drop. On a Raspberry Pi 5 with an AI HAT+ NPU, PINE sustains real-time enhancement across resolutions (25.8 FPS at 1080p, 17.6 FPS at QHD, and 8.1 FPS at 4K) with resolution-invariant NPU latency (6.6-6.7 ms). Controlled-turbidity water-tank experiments on unseen scenes verify stable end-to-end operation at 15.6-16.4 FPS across 0.35-10.3 NTU without test-time adaptation. Code and models are available at https://github.com/Jinxinshao/PINE.