RPL-UIE: Reference-Guided Prior Learning for Content-Preserving Underwater Image Enhancement
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: We're looking at the paper: "RPL-UIE: Reference-Guided Prior Learning for Content-Preserving Underwater Image Enhancement".
Dev: Here's the gist of the paper. Underwater image enhancement (UIE) aims to recover clear images from observations affected by wavelength-dependent absorption, scattering, and spatially nonuniform degradation To address this issue, RPL-UIE proposes a two-stage teacher–student framework for reliable prior learning This framework involves a teacher stage where the network learns reliable and complementary spatial priors from paired degraded and reference images In the student stage, the network learns to emulate the teacher’s prior extraction capability, providing more reliable restoration guidance without requiring reference images at inference To reduce the prior-learning discrepancy between models, Residual Prior Refinement Diffusion (RPRD) and Frequency-Aware Prior Residual Calibration (FPRC) are developed Framework Overview The RPL-UIE framework is structured in two stages: Stage I focuses on teacher prior learning, and Stage II involves student prior estimation and calibration.
Rosa: In Stage I, the Appearance Spatial Prior Encoder (ASPE) and Photometric Spatial Prior Encoder (PSPE) learn complementary teacher spatial priors from paired degraded and reference images.
Dev: The appearance teacher prior captures color, contrast, and texture visibility The photometric teacher prior characterizes reflectance–illumination properties related to color attenuation and brightness variation.
Rosa: These teacher priors guide image reconstruction in Stage I and provide reliable supervision targets for the student stage.
Paper summary: Rosa: They set up this teacher stage where they have two encoders, one called ASPE for appearance and one called PSPE for photometry. These encoders learn complementary priors from paired degraded and reference images. Dev The appearance teacher prior captures things like color, contrast, and texture visibility in the image. Rosa And the photometric teacher prior focuses on properties related to how light hits the object, like color attenuation and brightness variation. Dev Those teacher priors are supposed to guide image reconstruction in that first stage and give reliable supervision targets for what comes next.
Taro: I wonder if learning those priors from pairs of images is really going to capture everything needed for real-world stuff when the degradation is totally unpredictable. Rosa That's a good point, Taro, because they then move to the student stage where the student estimates its own coarse priors just from the degraded input itself. Dev The appearance part of that student, ASPE, estimates a spatial coarse prior that mainly characterizes color distribution and local contrast in the degraded image. Taro So it's trying to guess what the appearance looked like before it got underwater.
Rosa: And then for photometry, PSPE estimates a photometric spatial coarse prior focusing on spatial variations in color attenuation and illumination visibility within that same degraded input. Dev They collect these two coarse priors as a set of appearance and photometric coarse priors, which feeds into the refinement mechanisms. Taro But how do they handle the gap between what the student guesses and what the teacher actually knows?
Rosa: That’s where they introduce Residual Prior Refinement Diffusion, or RPRD. They model prior refinement as a residual diffusion process moving from those student coarse priors toward their teacher counterparts. Dev RPRD uses that coarse prior as an anchor and only learns the residual correction needed to get to the teacher's version. Taro That sounds like it focuses on fixing the small errors rather than learning everything from scratch.
Rosa: Then they have Frequency-Aware Prior Residual Calibration, or FPRC. This calibrates those residuals between each refined prior and its coarse counterpart by keeping stable low frequency residual components and adaptively adjusting the high frequency detail residuals based on their reliability. Dev So it’s like a way to decide which parts of the correction are reliable enough to apply fully without introducing noise.
Taro: It makes sense that they'd want to isolate those reliable low-frequency components, because in underwater images, you usually have a lot of smooth color shifts across large areas. Rosa So the whole point is that these calibrated priors are injected into the reconstruction network via Spatial Prior Modulation, or SPM. Dev SPM converts those spatial priors into a multiplicative conditioning signal for the reconstruction features. Taro So it’s using these learned corrections to steer the enhancement process itself.
Rosa: The final enhanced image is generated by feeding that everything into G theta with the degraded input, and then those three priors—the appearance, photometric, and the refined ones—as conditioning signals. Dev It’s a structured way of guiding the restoration network instead of letting it guess everything randomly based on poor input data.
Conclusion: Rosa: So to wrap up this paper "RPL-UIE: Reference-Guided Prior Learning for Content-Preserving Underwater Image Enhancement," they developed this two stage teacher student framework for reliable prior learning. Dev They claim that by having the student emulate the teacher’s prior extraction capability, the enhancement guidance becomes much more dependable even without reference images at inference time. Rosa The authors found competitive restoration performance across several underwater image enhancement benchmarks, achieving a PSNR of twenty-six point eight five dB and an SSIM of zero point nine three one on the UIEB dataset. Dev They also showed that downstream tasks like underwater object detection and instance segmentation got better results, increasing mAP50 and mAP50:ninety-five from sixty-three point zero to sixty-five point four and twenty-seven point four respectively on the UDD dataset.
Taro: What this means for us is that we have a systematic way to build priors that are robust because they are learned from paired data, not just inferred from noisy, single degraded images at runtime. Rosa It’s about making the prior learning process itself reliable first, and then using those reliable learned corrections to guide the actual image enhancement. Dev The ablation studies confirmed this; adding either appearance or photometric prior helped PSNR and LPIPS, but the full model with both RPRD and FPRC gave them the best overall balance in terms of color consistency, foreground visibility, and detail preservation.
Rosa: The title itself says "Reference-Guided Prior Learning." It highlights that you need some kind of reference or paired data to get a high quality result, especially when dealing with those complex underwater degradation issues like scattering and absorption. Dev It implies that for applications in marine observation or robotics, having access to some form of paired data isn't just nice; it's essential for building a trustworthy system. Taro The limitation they pointed out is that while the performance is strong on the tested benchmarks, they need to look into more extreme openworld conditions and perhaps weakly supervised adaptation because their current robustness in those tougher scenarios needs more testing.
Rosa: It’s a solid piece of work because it shows how structuring the learning process—separating teacher prior learning from student estimation—can lead to better performance than just throwing a big generative model at the degraded image alone. Dev It moves the focus from just generating an image to systematically building a reliable guide for restoration.
College of Future Information Technology, Fudan University · Institute of Artificial Intelligence (TeleAI), China Telecom, China · Faculty of Robot Science and Engineering, Northeastern University, Shenyang, China · School of Artificial Intelligence, Optics and Electronics (iOPEN), Northwestern Polytechnical University, Xi’an
eess.IV, cs.CV
Submitted: 2026-07-31
Updated: 2026-11-05
Comments: 19 pages, 10 figures
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 100/100
The gist: Underwater image enhancement (UIE) aims to recover clear images from observations affected by wavelength-dependent absorption, scattering…
Key concepts
- Teacher–Student Framework
- A two-stage approach where a 'teacher' network learns reliable spatial priors from paired degraded and reference images. The 'student' network then learns to mimic the teacher's prior extraction ability without needing reference images during actual image enhancement, ensuring more consistent guidance.
- Appearance Spatial Prior Encoder (ASPE)
- This encoder is part of the teacher stage, designed to learn spatial priors related to how colors, contrast, and textures are visible in an image. It captures the visual characteristics necessary for color and texture restoration.
- Residual Prior Refinement Diffusion (RPRD)
- RPRD models prior refinement as a diffusion process. It uses the student's coarse prior as a starting point and learns only the residual correction needed to match the teacher's refined prior, effectively bridging the gap between different model estimations.
- Spatial Prior Modulation (SPM)
- This mechanism takes the learned spatial priors (appearance and photometric) and converts them into a multiplicative conditioning signal. This signal is then injected into a reconstruction network to guide it in enhancing the degraded image toward a clearer result.
Terminology
Summary
Underwater image enhancement (UIE) aims to recover clear images from observations affected by wavelength-dependent absorption, scattering, and spatially nonuniform degradation To address this issue, RPL-UIE proposes a two-stage teacher–student framework for reliable prior learning This framework involves a teacher stage where the network learns reliable and complementary spatial priors from paired degraded and reference images In the student stage, the network learns to emulate the teacher’s prior extraction capability, providing more reliable restoration guidance without requiring reference images at inference To reduce the prior-learning discrepancy between models, Residual Prior Refinement Diffusion (RPRD) and Frequency-Aware Prior Residual Calibration (FPRC) are developed
Framework Overview
The RPL-UIE framework is structured in two stages: Stage I focuses on teacher prior learning, and Stage II involves student prior estimation and calibration. In Stage I, the Appearance Spatial Prior Encoder (ASPE) and Photometric Spatial Prior Encoder (PSPE) learn complementary teacher spatial priors from paired degraded and reference images. p. 7 The appearance teacher prior captures color, contrast, and texture visibility The photometric teacher prior characterizes reflectance–illumination properties related to color attenuation and brightness variation. p. 7 These teacher priors guide image reconstruction in Stage I and provide reliable supervision targets for the student stage. p. 2
Student Prior Estimation
In Stage II, the student pathway estimates the corresponding coarse priors solely from the degraded input. ASPE first estimates the appearance spatial coarse prior, which primarily characterizes color distribution, local contrast, and texture visibility of the degraded image. p. 9 Using the degraded photometric representation, PSPE estimates the corresponding photometric spatial coarse prior emphasizing spatial variations in color attenuation, illumination, and reflectance visibility The complementary spatial coarse priors are collected as a set of appearance and photometric coarse priors. p. 9
Prior Refinement Mechanism
Residual Prior Refinement Diffusion (RPRD) is introduced to reduce the prior-learning discrepancy by modeling prior refinement as residual diffusion from student-estimated coarse priors toward their teacher counterparts. RPRD uses the coarse prior as an anchor and learns only the residual correction toward its teacher counterpart Frequency-Aware Prior Residual Calibration (FPRC) then calibrates the residual between each refined prior and its coarse counterpart by retaining stable low-frequency residual components and adaptively modulating high-frequency detail residuals according to their reliability. p. 7
Reconstruction Guidance
The resulting calibrated priors are injected into the reconstruction network through Spatial Prior Modulation (SPM) to guide image enhancement. SPM converts the corresponding spatial prior into a multiplicative conditioning signal for the reconstruction features The final enhanced image is generated as ˆI = Gθ(Id, P˜pho, P˜app). p. 4
Experimental Validation
Extensive experiments on multiple UIE benchmarks and downstream underwater object detection and instance segmentation tasks demonstrate competitive restoration performance RPL-UIE achieved the best PSNR, SSIM, LPIPS, and UCIQE values of 26.85 dB, 0.931, 0.081, and 4.359 in Table 1 on UIEB. p. 5 Downstream detection experiments showed that RPL-UIE increased mAP50 and mAP50:95 from 63. p. 10 and 25.4 to 65.4 and 27.4 on the UDD dataset. Real-world ROV experiments further supported the robustness of RPL-UIE for underwater visual perception. p. 1
Ablation Analysis
Ablation studies verified the effectiveness of the proposed components, showing that adding either appearance or photometric prior improved PSNR and LPIPS. Introducing RPRD yielded further gains, and FPRC enabled the Full model to achieve the best PSNR, SSIM, and LPIPS. p. 0 The Full model with FPRC achieved the most balanced overall result in color consistency, foreground visibility, and detail preservation. p. 6
Limitations
Although RPL-UIE demonstrates consistent performance across multiple benchmark datasets and degradation conditions, its robustness under more extreme openworld underwater conditions warrants further investigation. Future work will explore larger and more diverse real-world underwater datasets, weakly supervised adaptation, and degradation-aware data construction to improve robustness. p. 1
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Robust reference-free inference in Stage II allows for image enhancement without requiring reference images at inference by enabling a student network to
emulate the teacher’s prior extraction capability.
-
The use of Residual Prior Refinement Diffusion (RPRD) reduces the
prior-learning discrepancy between the teacher and student models
by usingthe coarse priors as anchors and progressively predicts the necessary corrections in the residual space.
-
Frequency-Aware Prior Residual Calibration (FPRC) ensures reliable guidance for image reconstruction by retaining stable low-frequency residual components while selectively modulating high-frequency detail residuals
according to their spatial-channel reliability.
-
The proposed RPL-UIE framework can provide
more reliable restoration guidance for the enhancement process without requiring reference images at inference.
-
Downstream underwater object detection and instance segmentation experiments demonstrate that enhanced images support
improved utility for visual perception,
leading to higher metrics like mAP50 and AP50:95. -
Real-world ROV experiments show that RPL-UIE improves
local structural visibility and interframe feature matching in the captured turbid underwater scene,
which benefits subsequent underwater visual navigation.
Related papers
- HDNet: High-resolution Dual-domain Learning for Spectral Compressive Imaging
- Mask-guided Spectral-wise Transformer for Efficient Hyperspectral Image Reconstruction
- FDIR: Fidelity-Anchored Decompressed Image Restoration for Human and Machine Perception
- A Disease-Centric Vision-Language Foundation Model for Precision Oncology in Kidney Cancer
- CADS: A Comprehensive Anatomical Dataset and Segmentation for Whole-Body Anatomy in Computed Tomography
- Task-Aware Bitrate Thresholds for Codec-Robust Ultrasound Segmentation