RPL-UIE's teacher-student AI enhances underwater images with reliable priors
This two-stage framework restores clear underwater images without needing reference photos at inference.
Underwater images suffer from wavelength-dependent absorption, scattering, and nonuniform degradation, and existing generative methods can lose critical semantics. In response, Yifan Chen, Jiaming Liu, Ye Zheng, Zhe Sun, and Tao Chen from Fudan University propose RPL-UIE, a two-stage teacher-student framework designed to learn reliable priors for underwater image enhancement (UIE). The teacher network extracts complementary spatial priors—characterizing appearance and photometric properties—from paired degraded and reference images. At inference, the student network takes only the degraded image as input and emulates the teacher's prior extraction ability, producing restoration guidance without needing any reference image. This architecture directly addresses the semantic drift problem common in generative UIE methods.
To close the gap between teacher and student, the authors introduce two key innovations: Residual Prior Refinement Diffusion (RPRD) and Frequency-Aware Prior Residual Calibration (FPRC). RPRD treats coarse priors as anchors and progressively predicts corrections in residual space, stabilizing the learning process. FPRC preserves low-frequency residual components while selectively modulating high-frequency details, yielding calibrated priors for high-quality reconstruction. Experiments across multiple UIE benchmarks show competitive restoration performance. Critically, downstream tests on underwater object detection and instance segmentation confirm that images enhanced by RPL-UIE significantly improve visual perception accuracy. The method was also validated on real-world data captured by a remotely operated vehicle (ROV), underscoring its practical value for marine exploration, inspection, and autonomous underwater navigation.
- RPL-UIE uses a two-stage teacher-student framework, eliminating the need for reference images at inference time
- Introduces RPRD (diffusion-based residual refinement) and FPRC (frequency-aware calibration) to reduce prior-learning discrepancy
- Outperforms on UIE benchmarks and boosts downstream underwater object detection and instance segmentation, with real-world ROV validation
Why It Matters
Clearer underwater imagery from RPL-UIE boosts ROV and marine robot perception, improving exploration, inspection, and autonomous navigation reliability.