A Generative Data Framework with Reliable Supervision for Underwater Image Restoration and Enhancement

arXiv:2511.14521 · cs.CV · Submitted 2025-11-18 · 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: "A Generative Data Framework with Reliable Supervision for Underwater Image Restoration and Enhancement".

Jane: Here's the gist of the paper. Underwater image restoration and enhancement are crucial for correcting color distortion and restoring image details, thereby establishing a fundamental basis for subsequent underwater visual tasks This research proposes using in-air natural images as unambiguous reference targets and translating them into underwater-degraded versions, thereby constructing synthetic datasets that provide authentic supervision signals for model learning Extensive quantitative and qualitative experiments across 6 representative network architectures and 3 independent test sets show that models trained on our synthetic data achieve comparable or superior color restoration and generalization performance to those trained on existing benchmarks This research provides a reliable and scalable data-driven solution for underwater image restoration and enhancement Pruning redundant reasoning tokens The authors address the scarcity of high-quality paired datasets by proposing a generative data framework based on unpaired image-to-image translation to create synthetic datasets with precise ground-truth labels This framework constructs synthetic datasets with precise ground-truth labels, which facilitate the learning of an accurate mapping from degraded underwater images to their pristine scene appearances The core idea is to transform natural in-air images with high color fidelity into visually authentic underwater-degraded counterparts This process yields a paired dataset: Dsynth = (x, yˆd) where yˆd = G(x; d) How the training signal is built The training signal is built by employing an advanced I2IT model to efficiently transform in-air images sourced from the RAISE, ImageNet, and iNaturalist-12K datasets into 6 representative categories of degraded underwater images This methodology results in a large and diverse set of paired samples consisting of realistically degraded images and their corresponding clear references, thus providing a robust foundation for supervised learning Specifically, the framework can be formulated as a mapping function G that transforms a source in-air image x ∈ X into a degraded underwater image yˆ ∈ Y conditioned on a specific degradation type d, where d ∈ (Blue, Low-Light, Deep Blue, Deep Green, Green, and Blurry) Where the gains come from The gains come from leveraging existing public datasets to efficiently generate diverse high-quality paired samples while eliminating the need for labor-intensive collection of strictly paired crossdomain imagery The authors construct and release 2 large-scale synthetic datasets, UWNature and UWImgNetSD, each comprising 10,000 pairs designed to encompass 6 diverse underwater degradation types The main contributions include proposing a novel generative data framework that solves core issues of existing datasets by transforming natural images with explicit color fidelity into underwater degraded images Quantitative validation and practical advantages The effectiveness is validated by evaluating models trained on our data and open datasets across 6 network architectures and 3 independent test sets, demonstrating superior performance in color restoration, visual naturalness, and cross-domain generalization Models trained on our synthetic datasets consistently achieve superior performance in color restoration, visual naturalness, and cross-domain generalization compared to those trained on existing public datasets Furthermore, the method is fully automated and scalable, eliminating the need for physical underwater reference collection and errorprone manual ground truth selection The application confirms that models trained on our synthetic data can tackle complex real-world underwater degradation, proving its practical transferability Ablation study confirmation The ablation study confirms that diverse degradation types are crucial for synthetic dataset construction, as mixed-type datasets enhance model generalization and robustness across underwater scenarios by preventing overfitting Models trained on mixed-type datasets achieve significantly higher UCIQE scores than those trained on any single-type dataset, indicating that having diverse degradation types is crucial for synthetic dataset construction The ablation study confirms that diverse degradation types are crucial for synthetic dataset construction, as mixed-type datasets enhance model generalization and robustness across underwater scenarios by preventing overfitting Application validation The application confirms the effectiveness of synthetic data-driven models on standard benchmarks by restoring color casts in real-world environments using the DMunderwater model trained on our UWNature dataset The method effectively corrects color casts and restores the colorboard’s appearance, demonstrating a significant correction of the aquatic color distortion This experiment confirms that models trained on our synthetic data can tackle complex real-world underwater degradation, proving its practical transferability This paper addresses the critical data bottleneck in underwater image restoration and enhancement through a generative data framework Future work will explore integrating vision-language models for semantic-aware enhancement and jointly optimizing the data framework with downstream tasks like object detection and segmentation The authors conclude that leveraging advanced generative models for dataset construction is a highly effective and scalable path forward for underwater image restoration and enhancement REFERENCES

Title and authors: Tom: So let's talk about what they actually propose in this paper. The title itself suggests they are building a system for restoration and enhancement using synthetic data as the backbone.

Jane: They’re proposing this generative data framework, which is essentially creating a way to get reliable supervision signals for model learning by transforming natural images into underwater-degraded versions with precise ground truth labels.

Lu: The main improvement they highlight is that they move away from relying on manually selected results from existing enhancement algorithms, which often lack globally consistent color and authentic supervision, and instead use high-quality natural scenes as the source.

Meng: What I like is that they are using data from RAISE, ImageNet, and iNaturalist-12K to generate these degraded samples across six different categories of degradation types: Blue, Low-Light, Deep Blue, Deep Green, Green and Blurry > <ref:2511.14521#pg2>

Lalam: It results in a large and diverse set of paired samples consisting of realistically degraded images matched with their corresponding clear references. This provides a robust foundation for supervised learning because the color information is accurate and free from ambiguity >

The paper's summary: Tom: They summarize this by saying that this method overcomes the limitations of existing datasets in terms of color authenticity and establishes a reliable data foundation for improving both restoration accuracy and cross-scene generalization capability.

Jane: So, to put it simply, they are showing that by translating natural images into underwater degraded versions, you can build a dataset that helps models learn how to map those degraded pictures back to their original pristine appearances.

Lu: That mapping function G transforms the source in-air image x into a degraded underwater image y hat d conditioned on a specific degradation type d, where d is one of those six types >

Meng: It’s about scaling up the data generation process rapidly. The paper emphasizes that the translation process itself facilitates rapid and scalable data generation, which is good for getting enough samples quickly.

Lalam: They construct two large-scale synthetic datasets called UWNature and UWImgNetSD, each with ten thousand pairs designed to cover those six diverse underwater degradation types >

The paper's improvements: Tom: The paper points out a few key improvements over what’s already out there. First, they show that models trained on their synthetic data achieve comparable or even superior color restoration and generalization performance when tested across six representative network architectures and three independent test sets >

Jane: That’s significant because it proves that the synthetic data isn't just good for one model type; it works well across different ways of building a deep learning model.

Lu: And they also confirm that using mixed-type datasets, meaning you have samples with multiple degradation forms, enhances model generalization and robustness because it helps prevent overfitting >

Meng: I saw the ablation study mentioned where they compare single-type datasets against mixed-type ones, and the mixed-type ones showed significantly higher UCIQE scores, which means having diversity in those degradation types really helps stability.

Lalam: They also show that this method is fully automated and scalable because it eliminates the need for labor-intensive collection of physical underwater reference material and errorprone manual ground truth selection >

Conclusion: Tom: So to wrap up, the paper "A Generative Data Framework with Reliable Supervision for Underwater Image Restoration and Enhancement" basically argues that using generative data frameworks to create synthetic datasets with authentic supervision signals is a reliable way forward.

Jane: It suggests that by focusing on translating natural images into underwater degraded versions, we get a scalable way to build better training data than what’s currently available from manual curation.

Lu: The implication for the field is that this provides a reliable data foundation for improving color restoration accuracy and cross-scene generalization capability in underwater vision tasks >

Meng: From an engineering standpoint, it means we can deploy models trained on this synthetic data that are more robust to various real-world conditions, provided the translation model G is accurate enough.

Lalam: The authors conclude that leveraging advanced generative models for dataset construction is a highly effective and scalable path forward for underwater image restoration and enhancement >

Institute of Artificial Intelligence (TeleAI) · China Telecom Shanghai Institute of Artificial Intelligence (TeleAI) · Department of Automation Shanghai Jiao Tong University Department of Automation

cs.CV

Submitted: 2025-11-18

Updated: 2026-11-10

Comments: Substantially revised version with expanded analyses of synthetic supervision reliability, including appearance distributions, structural consistency, depth-aware modulation, perceptual evaluation, and restoration behavior

Code: https://github.com/yftian2025/SynUIEDatasets

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

Importance score: 91/100

The gist: Underwater image restoration and enhancement are crucial for correcting color distortion and restoring image details…

Key concepts

Generative Data Framework
This is a method that uses unpaired image-to-image translation to create new training data. It takes clear, in-air photos and transforms them into realistic underwater versions, generating large sets of paired images (degraded image and its clear reference) without needing manual collection.
Image-to-Image Translation (I2IT)
An advanced model used to perform the core transformation. It learns the mapping function G that converts a source in-air image into a degraded underwater image, conditioned on specific degradation types like blue or low light. This allows for automated creation of synthetic training pairs.
Synthetic Datasets (UWNature/UWImgNetSD)
These are large-scale datasets created by the framework, comprising 10,000 pairs each. They consist of realistic underwater images paired with their pristine counterparts derived from natural images. These datasets provide authentic supervision signals for training restoration models.
Mixed-Type Datasets
Using synthetic data containing multiple degradation types (e.g., blue, blurry, low-light) instead of single-type datasets improves model robustness. This diversity prevents overfitting and ensures the model can handle a wider variety of complex underwater scenarios effectively.

Terminology

Summary

Underwater image restoration and enhancement are crucial for correcting color distortion and restoring image details, thereby establishing a fundamental basis for subsequent underwater visual tasks This research proposes using in-air natural images as unambiguous reference targets and translating them into underwater-degraded versions, thereby constructing synthetic datasets that provide authentic supervision signals for model learning Extensive quantitative and qualitative experiments across 6 representative network architectures and 3 independent test sets show that models trained on our synthetic data achieve comparable or superior color restoration and generalization performance to those trained on existing benchmarks This research provides a reliable and scalable data-driven solution for underwater image restoration and enhancement

Pruning redundant reasoning tokens

The authors address the scarcity of high-quality paired datasets by proposing a generative data framework based on unpaired image-to-image translation to create synthetic datasets with precise ground-truth labels This framework constructs synthetic datasets with precise ground-truth labels, which facilitate the learning of an accurate mapping from degraded underwater images to their pristine scene appearances The core idea is to transform natural in-air images with high color fidelity into visually authentic underwater-degraded counterparts This process yields a paired dataset: Dsynth = (x, yˆd) where yˆd = G(x; d)

How the training signal is built

The training signal is built by employing an advanced I2IT model to efficiently transform in-air images sourced from the RAISE, ImageNet, and iNaturalist-12K datasets into 6 representative categories of degraded underwater images This methodology results in a large and diverse set of paired samples consisting of realistically degraded images and their corresponding clear references, thus providing a robust foundation for supervised learning Specifically, the framework can be formulated as a mapping function G that transforms a source in-air image x ∈ X into a degraded underwater image yˆ ∈ Y conditioned on a specific degradation type d, where d ∈ (Blue, Low-Light, Deep Blue, Deep Green, Green, and Blurry)

Where the gains come from

The gains come from leveraging existing public datasets to efficiently generate diverse high-quality paired samples while eliminating the need for labor-intensive collection of strictly paired crossdomain imagery The authors construct and release 2 large-scale synthetic datasets, UWNature and UWImgNetSD, each comprising 10,000 pairs designed to encompass 6 diverse underwater degradation types The main contributions include proposing a novel generative data framework that solves core issues of existing datasets by transforming natural images with explicit color fidelity into underwater degraded images

Quantitative validation and practical advantages

The effectiveness is validated by evaluating models trained on our data and open datasets across 6 network architectures and 3 independent test sets, demonstrating superior performance in color restoration, visual naturalness, and cross-domain generalization Models trained on our synthetic datasets consistently achieve superior performance in color restoration, visual naturalness, and cross-domain generalization compared to those trained on existing public datasets Furthermore, the method is fully automated and scalable, eliminating the need for physical underwater reference collection and errorprone manual ground truth selection The application confirms that models trained on our synthetic data can tackle complex real-world underwater degradation, proving its practical transferability

Ablation study confirmation

The ablation study confirms that diverse degradation types are crucial for synthetic dataset construction, as mixed-type datasets enhance model generalization and robustness across underwater scenarios by preventing overfitting Models trained on mixed-type datasets achieve significantly higher UCIQE scores than those trained on any single-type dataset, indicating that having diverse degradation types is crucial for synthetic dataset construction The ablation study confirms that diverse degradation types are crucial for synthetic dataset construction, as mixed-type datasets enhance model generalization and robustness across underwater scenarios by preventing overfitting

Application validation

The application confirms the effectiveness of synthetic data-driven models on standard benchmarks by restoring color casts in real-world environments using the DMunderwater model trained on our UWNature dataset The method effectively corrects color casts and restores the colorboard’s appearance, demonstrating a significant correction of the aquatic color distortion This experiment confirms that models trained on our synthetic data can tackle complex real-world underwater degradation, proving its practical transferability

Conclusion and future work

This paper addresses the critical data bottleneck in underwater image restoration and enhancement through a generative data framework Future work will explore integrating vision-language models for semantic-aware enhancement and jointly optimizing the data framework with downstream tasks like object detection and segmentation The authors conclude that leveraging advanced generative models for dataset construction is a highly effective and scalable path forward for underwater image restoration and enhancement

REFERENCES

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. Data-driven restoration using synthetic datasets that provide authentic supervision signals for model learning. This allows models to learn a mapping from degraded underwater images to their pristine scene appearances with precise ground-truth labels.

  2. Construction of large-scale, pixel-level aligned training pairs through an image translation framework. This enables the generation of a dataset where the reference images for this dataset originate from real natural images.

  3. Training state-of-the-art enhancement algorithms on synthetic and existing public datasets to compare performance across different network architectures. This validates the universality of synthetic data across network architectures and optimization paradigms.

  4. Achieving superior color restoration, visual naturalness, and cross-domain generalization compared to models trained on existing public datasets. Models can correct common underwater blue/green color casts with high chroma standard deviation because the reference images are color-rich, unbiased natural or synthetic images.

  5. Robustness across different model architectures by training on mixed-type synthetic datasets that contain multiple degradation forms. This prevents overfitting and ensures stable and effective restoration across various degradation types.

  6. Real-world validation of the DMunderwater model using the UWNature dataset to correct color casts and restore the colorboard's appearance in blue-cast clear water and turbid, low-light water environments.

Sources

Related papers