RQUL-UIE: Revitalizing Quality-Unstable Labels for Underwater Image Enhancement via In-Dataset Self-Supervision

arXiv:2606.06176 · cs.CV · Submitted 2026-06-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "RQUL-UIE: Revitalizing Quality-Unstable Labels for Underwater Image Enhancement via In-Dataset Self-Supervision".

Paper summary: Tom: So, to wrap up this paper, RQUL-UIE is a diffusion framework that tackles label noise by treating it like a level-wise denoising problem >

Jane: It suggests that using quality scores derived from another pre-trained diffusion model, and then mapping those scores to specific denoising steps, allows the system to use noisy labels effectively >

Lu: The key is reformulating the rigid pairwise supervision into this process so it can revitalize those labels during training without compromising the dataset's structural diversity >

Meng: For deployment at scale, we need to watch how reliably those noise-level indices work when faced with entirely new types of underwater degradation not seen in the training data >

Tom: The authors confirm that this strategy successfully exploits in-dataset re-learning opportunities and that the Fourier refinement module helps fix those high-frequency details that usually get lost >

Jane: It really changes how we think about noisy labels; it moves away from just discarding them or using them blindly, toward a process where the model learns to trust the label quality itself >

Conclusion: Tom: So, we're wrapping up on RQUL-UIE, which is about taking those noisy labels in underwater image enhancement and figuring out how to use them better by treating it like a denoising process that happens right inside the training set.

Jane: Right, so the core idea is using a diffusion model to check if a label is good or bad in a way that doesn't need new training data for every single label. It uses semantic embeddings from an already trained diffusion model to get those quality scores.

Lu: What I find really interesting about the authors is how they take those continuous quality scores and turn them into discrete noise levels, which then guides a multi-step denoising process. They’re essentially telling the model exactly how many steps it needs to take for each label during training.

Meng: From an engineering standpoint, that sounds smart because you don't have to filter out the bad labels entirely, which is what usually causes issues with dataset diversity. But I wonder about the computational cost of running this multi-step process for every single label in every batch.

Tom: That's a valid concern, Meng. The paper shows they’ve kept it manageable by incorporating a Fourier-based refinement network to fix those high-frequency details that usually get messed up by low-quality labels. They claim the total computational cost stays quite low compared to standard diffusion pipelines on this resolution.

Jane: It sounds like they aren't just fixing the label problem, they're also enhancing the visual quality of the output because of that texture refinement part. The results show a clear improvement in restoration quality and visual fidelity compared to other methods.

Lu: The authors prove that discarding low-quality labels directly actually hurts performance on some tasks, so this level-wise strategy is definitely better than just throwing away the bad ones. It seems to exploit the latent intelligence of the whole dataset instead of just relying on perfect pairs.

Meng: So, it's not a magic fix, but a smarter way to use what you already have without sacrificing data variety or paying for massive retraining cycles. But how robust is this noise-level ranking when you move to genuinely different underwater conditions?

Tom: That’s the next big question, Meng. The authors show strong results on their test set, but we need to see how it holds up when the water physics completely change from what they trained on.

Jane: Exactly. So, RQUL-UIE gives us a framework that lets us leverage our existing underwater data much more effectively by being smarter about which labels we trust during the training process.

The Hong Kong Polytechnic University

cs.CV

Submitted: 2026-06-04

Updated: 2026-11-17

Code: https://github.com/Haru2022/RQUL-UIE

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 91/100

The gist: Underwater Image Enhancement (UIE) is essential for mitigating degradations caused by water medium Page 1</ref:2…

Key concepts

Label Quality Instability (Di)
This measures how much a raw label pair differs from each other using semantic perception embeddings from a pre-trained diffusion model. A higher instability score indicates the label is less reliable or noisier. This metric is used to determine the necessary number of denoising steps required to transform an underwater image into that specific label.
Quantized Noisy Level ($t_i$)
The continuous instability score ($D_i$) is mapped into discrete 'noisy levels' ($t_i$). This quantization represents the exact number of denoising steps needed for an image to become a valid label. By using these quantized levels, the training process can be guided to appropriately supervise labels at different stages.
Level-wise Diffusion Denoising
Instead of treating all labels equally, this strategy reformulates the learning task into a level-wise diffusion denoising process. This means that during training, images are denoised recurrently based on the specific quality level assigned to their corresponding label, ensuring low-quality labels are used constructively rather than damaging the model.
Fourier-based Texture Refinement Network (TRFDM)
This module is added to reconstruct high-frequency details that might be lost during low-dimension learning. It fuses local texture information from raw images using frequency domain modulation. This ensures that the final enhanced image has sharp, consistent textures, improving overall visual fidelity.

Terminology

Summary

Underwater Image Enhancement (UIE) is essential for mitigating degradations caused by water medium Page 1</ref:2, although learning-based methods often rely on paired datasets with unstable label quality which bottlenecks model performance Page 1</ref:3. p. 0 This paper proposes a diffusion-based, in-dataset self-supervised learning strategy designed to exploit the quality distribution of training labels Page 1</ref:4, specifically by evaluating label quality via semantic perception embeddings from a pre-trained diffusion model in a training-free manner Page 2</ref:5. p. 0 These quality scores are subsequently quantized into noise-level indices, guiding a multistep denoising process for level-wise supervision Page 2</ref:5, which prevents low-quality labels from degrading the model while maximizing their utility during training Page 2</ref:5, and a Fourier-based refinement network is incorporated to explicitly reconstruct high-frequency components Page 2</ref:5. p. 0 Extensive evaluations demonstrate that our method consistently outperforms SOTA approaches in restoration quality Page 1</ref:4. p. 0

Pruning redundant reasoning tokens

The authors shift the conventional supervised UIE paradigm under quality-unstable labels into an in-dataset selfsupervised framework by reformulating rigid pairwise mapping into a level-wise diffusion denoising process Page 2</ref:5. This strategy revitalizes noisy labels at appropriate supervisory stages without compromising dataset diversity Page 2</ref:5, and it effectively exploits the latent, collective intelligence of the entire dataset while strictly preserving its structural diversity from being compromised by mass data filtering Page 2</ref:5. p. 1

How the training signal is built

The label quality is reassessed and quantized in a training-free manner based on pre-trained diffusion prior, where the degree of label quality instability is quantized as different noisy levels Page 3</ref:5. Specifically, the degree of label quality instability is quantified as semantic difference between a raw-label pair Page 3</ref:5, defined as Di = z uw i − z la i squared Page 3</ref:5. p. 1 The quantized noisy level ti of the underwater image I uw I could be defined as the semantic difference ranking as t i = rankasc(Di)n · tmax Page 3</ref:5, which fundamentally quantifies the number of denoising steps required to transform I uw i into a label-like output Page 3</ref:5. p. 3 The final multi-step diffusion training loss is defined as Ldif f = z res i − z la i squared, where z res i is the denoised latent under the condition of z uw i = D(I uw i) with the queried denoising step tti = QLDB(I uw-I), meaning that the denoising process is performed recurrently tti times to obtain the final prediction Page 5</ref:5. p. 4

Where the gains come from

The Fourier-based texture refinement network, known as TRFDM, is designed to compensate for the high-frequency details inconsistency introduced by low-dimension learning Page 2</ref:5. This module fuses the inherent local texture details of the raw underwater images into the final output by frequency domain modulation Page 3</ref:5, and its computational overhead is virtually negligible compared to the standard Diffusion pipeline Page 13</ref:5. p. 7

Ablations About Discarding Low-Quality Labels

The necessity of LLSD is verified by showing that directly discarding low-quality labels is consistently inferior to the proposed level-wise learning strategy Page 8</ref:5. For instance, on UIEB, the average rank improves from 3. p. 575 for discard3 to 1.17 for vanilla Page 8</ref:5, indicating that filtering removes a large portion of training samples and harms distributional diversity Page 8</ref:5. p. 7

Computational Complexity Analysis (FLOPs) and Inference Efficiency

The total unified computational complexity under 512 × 512 resolution is 11.5307 GFLOPs, where Spatial Conv accounts for 8. p. 47955 GFLOPs and Pure Frequency Ops account for 2. p. 07352 GFLOPs Page 13</ref:5. p. 8 The TRFDM module introduces a mere 0. p. 429 M parameters and 11. p. 253 GFLOPs of computational workload, resulting in an average inference latency of 29.53 ms per frame Page 13. p. 8 This contrasts sharply with the standard Diffusion pipeline which demands ∼949. p. 257 M parameters and a protracted inference time of approximately 3. p. 70 seconds per generation Page 13. p. 8

Additional Enhanced Results for Quantitative Comparison

RQUL-UIE demonstrates a clear and massive margin of improvement compared to contemporary approaches, with the value of rankmetric dataset of the proposed RQUL-UIE being 1.83 Page 7</ref:5. p. 8 The qualitative comparison shows that RQUL-UIE exhibits a correct color pattern, high-contrast details, and overall visual fidelity compared to previous work Page 8</ref:5. p. 5 The superiority results from the designed level-wise denoising process where labels with various degrees of quality are properly used for the multi-step denoising so that the method could remove the global structure noise under diverse water conditions Page 8</ref:5. p. 5 The texture refinement module simultaneously recovers high-frequency details from the corresponding un-derwater images to improve image contrast Page 9</ref:5. p. 5

Conclusion

We have presented RQUL-UIE, a diffusion-based framework designed to address the challenge of label noise in underwater image enhancement Page 9</ref:5. Our work confirms that by reformulating rigid pairwise mapping into a level-wise denoising process, a pre-trained diffusion model can successfully revitalize quality-unstable labels for performance improvement Page 9</ref:5. p. 7 This strategy effectively exploits the possibility of in-dataset re-learning Page 9</ref:5, and our Fourierbased refinement module ensures high-frequency consistency, overcoming the inherent limitations of latent-space learning Page 9</ref:5. p. 7 Future research direction will target at optimizing the inference speed for real-time robotic applications Page 9</ref:5. p. 7

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

  1. Revitalize label quality in training by using a level-wise denoising process instead of direct supervision, which allows low-quality labels to mitigate the guidance of high-quality labels. This ensures that models learn from all data while preventing low-quality labels from degrading performance.

  2. Implement a Fourier based refinement network to explicitly reconstruct high frequency components and compensate for the high-frequency details inconsistency introduced by the Variational AutoEncoder (VAE) in low-dimension learning. This allows for the extraction of hidden texture details in raw underwater images to the final result by frequency domain modulation.

  3. Utilize a training free label quality reassessment based on semantic perception embeddings from a pre-trained diffusion model to quantize label instability into noise-level indices, which guide a multistep denoising process for level wise supervision. This mechanism prevents low quality labels from degrading the model while maximizing their utility during training.

  4. Design an inference pipeline that cascades Level wise Latent Space Denoising (LLSD) and Texture Refinement via Frequency Domain Modulation (TRFDM) to form a general pipeline for Underwater Image Enhancement, achieving superior restoration quality. This process ensures that the denoising step is set to maximum to remove visual degradations as much as possible during testing.

  5. Achieve high computational efficiency by introducing the TRFDM module, which has a parameter count of only 0.29 M and a total unified complexity of 11.53 GFLOPs, representing a fraction of additional burden compared to the standard diffusion pipeline's 949.57 M parameters and 3.0 seconds inference time. This allows for real-time robotic applications with a throughput of 33.86 Frames Per Second (FPS).

Sources

Related papers