PaperlayerView exact revision on arXiv

Source-linked AI summary

Rethinking Underwater Image Enhancement: Learning through Text-Guided Degradation Curriculum

Hongri Tian, Wenxue Li, Tian Ye, Sixiang Chen, Lujia Wang, Xianying Wang, Xinhu Zheng, Gang Hua, Lei Zhu

arXiv:2610.29569v1cs.CV

TL;DR

Underwater enhancement is limited by data that are scarce or fail to cover diverse real-world degradations. The paper combines text-guided degradation synthesis, progressive training, and transmission-guided one-step restoration; it reports state-of-the-art performance across multiple benchmarks.

  • Problem

    Underwater degradations vary across scenes and are difficult to model, while existing paired data are limited and do not adequately cover real-world diversity.

    Source: §1 · pp. 1–2

  • Method

    The framework synthesizes text-conditioned degradation variants, organizes them into a progressive curriculum, and uses a transmission-guided one-step diffusion enhancer.

    Source: §1 · pp. 2–3, §3.1.3 · p. 5

  • Results

    Across multiple underwater benchmarks, the method reports state-of-the-art performance and improvements across most restoration and reference-free metrics.

    Source: §1 · p. 3, §5 · p. 10

  • Takeaways & Limitations

    The paper concludes that controllable degradation modeling can provide structured supervision for robust underwater image enhancement.

    Source: §5 · p. 10

Abstract

from arXiv · show

Existing paired clean-degraded datasets have advanced learning-based underwater image enhancement, but they remain either limited in scale or difficult to acquire, and still fail to adequately cover the diverse degradations encountered in real-world underwater environments. Such scarce and insufficient data ultimately restrict the generalization of underwater image enhancement models, highlighting the need for realistic and controllable degradation modeling to better capture real underwater conditions and improve enhancement robustness. To this end, we propose a unified degradation-aware training framework for the underwater image enhancement task. A Text-Conditioned Degradation Augmentation Pipeline uses a text-guided diffusion model to synthesize diverse and semantically consistent degraded samples, providing paired data to support the subsequent curriculum learning strategy. The proposed Progressive Degradation Curriculum organizes training across multiple attributes and difficulty levels to promote balanced learning, and a Transmission-Guided One-Step Diffusion Model leverages this progressive process to better address the wide range of degradation severity and restoration ambiguity in underwater images. Extensive experiments on multiple underwater benchmarks demonstrate that our proposed method achieves superior performance. Our code and models will be released upon the publication of this work.

1 Introduction

Underwater degradations vary with scene conditions and are difficult to model, while paired training data are costly to acquire or insufficiently realistic. The paper combines text-controlled degradation synthesis, progressive curriculum learning, and transmission-guided one-step restoration.

1 Introduction

  • Underwater degradations vary with water type, depth, illumination, and particle density, making their distribution difficult to model and enhancement data-sensitive.The paper describes these factors as spatially coupled and scene-dependent.Source: §1 · pp. 1–2
  • Controlled tank capture and pseudo-reference selection are costly and difficult to scale, while simplified synthetic rendering often misses natural lighting and scattering.The paper identifies a need for realistic, controllable degradation modeling across varying degradation strengths and compositions.Source: §1 · p. 2
  • Text-guided diffusion synthesizes degradation variants conditioned on semantic cues, providing controllable data augmentation and difficulty cues for progressive enhancement training.The framework also injects a transmission prior into a one-step diffusion enhancer, reducing sampling cost relative to multi-step diffusion methods.Source: §1 · pp. 2–3
  • The authors report state-of-the-art performance on multiple underwater benchmarks.Their contributions also include a text-conditioned augmentation pipeline, progressive curriculum learning, and a transmission-guided one-step diffusion enhancer.Source: §1 · p. 3

2 Related Work

Prior underwater enhancement methods have advanced restoration, but their effectiveness remains constrained by training-data availability and diversity. Existing synthesis approaches leave a gap in controllable degradation modeling, which this work addresses with text-driven diffusion.

2.1 Underwater Image Enhancement

  • Prior-based methods depend on accurate water parameters and scene depth, while deep, adversarial, and diffusion approaches explore varied restoration cues and designs.The related work also notes that some adversarial and generative methods introduce instability or color over-saturation.Source: §2.1 · p. 3
  • Despite these advances, related underwater enhancement methods remain constrained by training-data availability and diversity.The paper presents this as an ongoing constraint across the reviewed methods.Source: §2.1 · p. 3

2.2 Underwater Image Synthesis

  • Physics-based synthesis can miss complex scattering, pseudo-references provide a fixed target without modeling forward degradation, and GAN methods can have domain bias and unstable training.These approaches therefore do not supply diverse, controllable degradation attributes such as water type and turbidity.Source: §2.2 · p. 3
  • Recent text- or depth-conditioned diffusion methods synthesize underwater scenes, whereas this work models controllable text-driven degradations.The TCDA pipeline links clean and degraded images with textual degradation prompts.Source: §2.2 · p. 4, Figure 2 · p. 4

3 Method

The method combines text-conditioned degradation synthesis with curriculum learning and transmission-guided one-step diffusion for underwater image enhancement. It uses controllable degradation prompts and a progressive training schedule to support restoration across varied degradation conditions.

3 Method

  • A two-stage framework synthesizes controllable degraded images, then uses progressive curriculum learning to train transmission-guided one-step restoration.The framework is designed for greater inference efficiency than multi-step diffusion methods.Source: §3 · p. 4

3.1 Text-Conditioned Degradation Augmentation

  • TCDA establishes a controllable mapping between clean images, degraded counterparts, and textual degradation prompts.The degradation diffusion model is trained using 12,483 degraded–clean pairs from four underwater datasets; only UIEB and LSUI training subsets are used for fair evaluation.Source: §3.1 · p. 4, §3.1.1 · p. 4

3.1.1 Degradation Captioning via MLLMs

  • Gemini captions describe six degradation attributes while excluding scene content, linking degradation cues to paired clean and degraded images.The attributes include turbidity, color-cast type and severity, suspended particles, lighting, haze or backscatter, and contrast reduction.Source: §3.1.1 · pp. 4–5

3.1.2 Flow-Based Degradation Modeling

  • Flow-based diffusion learns a text-conditioned vector field that transforms clean-image latents toward degraded-image latents under textual guidance.Clean and degraded images are encoded as latent representations, while frozen T5-XXL and CLIP-L encoders produce the text embedding; LoRA reduces training overhead.Source: §3.1.2 · p. 5

3.1.3 Controllable Degradation Synthesis

  • The prompt bank combines six degradation attributes with easy, moderate, or hard tiers to generate controllable degraded images from clean inputs.The attributes are turbidity, color cast, particles, lighting, haze or backscatter, and contrast.Source: §3.1.3 · p. 5

3.2 Progressive Degradation Curriculum Learning

  • The curriculum shifts from color and tone restoration toward scattering and contrast loss using progress-dependent sampling across three difficulty levels.The schedule uses a temperature factor of 5.0 in experiments; its evolving sampling probabilities are described as promoting progressive robustness.Source: §3.2 · pp. 5–7

3.3 Transmission-Guided One-Step Diffusion

  • TG-Diff conditions a single-step diffusion model on transmission maps concatenated with image latents and adds encoder-to-decoder skip features to preserve high-frequency detail.The transmission map is estimated with UDCP, and skip features are produced by a 1 × 1 convolution before being added to upsampled decoder features.Source: §3.3 · p. 7

4 Experiment

Experiments evaluate the method on paired and reference-free underwater benchmarks, where it achieves strong reported performance. Ablations examine TG-Diff components, curriculum learning, augmentation scale, and the controllability and distributional similarity of generated degradations.

4.1 Experimental Setup

  • Evaluation uses 790 UIEB training images, 100 test images, and 10,270 TCDA-augmented images divided among easy, moderate, and hard levels at a 5:5:3 ratio.Paired evaluation uses PSNR, SSIM, and LPIPS; unpaired evaluation uses UCIQE and UIQM, including on C60 and U45.Source: §4.1 · pp. 7–8

4.2 Comparison with State-of-the-Arts

  • TG-Diff achieves the strongest overall performance across UIEB, C60, and U45; on UIEB with TCDA, it records SSIM 0.927, PSNR 22.341, LPIPS 0.129, and UCIQE 0.613.On reference-free C60 and U45, it obtains the highest UCIQE and UIQM scores in most cases; TCDA improves most reported metrics across compared methods.Source: §4.2 · p. 8
  • Visual comparisons on UIEB report clearer visibility, more natural color balance, and better local detail for TG-Diff, with further fidelity gains from TCDA in heavily degraded scenes.TCDA-enhanced results are described as removing haze more thoroughly and recovering clearer textures and structures.Source: Figure 5 · p. 8, §4.2 · p. 9

4.3 Ablation Studies

  • In TG-Diff ablations, a trainable encoder improves feature adaptability, skip connections enhance structural fidelity, and the UDCP prior improves SSIM and LPIPS while maintaining comparable PSNR.On LSUI, TG-Diff is reported to outperform representative UIE networks across reference-based and non-reference metrics; curriculum learning yields the best overall performance among difficulty-sampling settings.Source: §4.3 · p. 9, Table 4 · p. 10
  • Increasing the augmented-to-original ratio from 1× to 3× and 5× generally improves restoration fidelity and perceptual quality.The Table 5 caption likewise characterizes increasing the synthetic-degradation ratio as improving overall enhancement quality.Source: §4.3 · p. 9, Table 5 · p. 10
  • Text prompts produce degradation transitions from mild color distortion to severe scattering and contrast loss, while UCIQE and UIQM decline consistently from easy to hard levels.The generated degradations have lower FID than UWGAN, UWCNN, and UWNR against real UIEB and LSUI test images; TG-Diff trained on TCDA data also outperforms training on UWCNN-generated degradations.Source: §4.3 · p. 10, Figure 8 · p. 10, Table 6 · p. 10

5 Conclusion

The paper presents controllable degradation modeling, curriculum learning, and transmission-guided one-step restoration as a framework for UIE. Experiments on multiple benchmarks show improvements across most restoration and reference-free metrics, while future work targets underwater video and broader real-world degradations.

5 Conclusion

  • The framework combines TCDA, a progressive degradation curriculum, and transmission-guided one-step TG-Diff to address scale and diversity limitations in paired UIE data.The conclusion characterizes these components together as structured supervision for robust UIE.Source: §5 · p. 10
  • Experiments show improvements across most restoration and reference-free metrics, and future work will extend the framework to temporally consistent underwater video and broader real-world degradation conditions.The paper states that the generated augmented training data is available at its project repository.Source: §5 · pp. 10–11

Related papers

Loading 2610.29569v1…