Research map / Image generation / Image restoration

Dehaze-RetinexGAN: Real-World Image Dehazing via Retinex-based Generative Adversarial Network

AAAI 2025 · Image generation › Image restoration · approach: Dehazing and deraining · 17 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Dehaze-RetinexGAN, a lightweight two-stage Retinex-based GAN that dehazes real images using unpaired data via self-supervised pre-training (DecomNet) and weakly-supervised adversarial fine-tuning with a dual DTCWT attention U-Net.

Problem

Real-world image dehazing: deep dehazing networks trained on paired synthetic data generalize poorly to real hazy scenes, and prior unpaired/real-world methods produce unnatural results or are slow.

Approach

Reduces dehazing to illumination-reflectance decomposition of the inverted hazy image (Retinex-dehazing duality), uses a self-supervised reconstruction loss tying the dehazed output back to the input hazy image, and adds a dual DTCWT-based channel/spatial attention U-Net fine-tuned with adversarial learning against unpaired clean images.

Key results

• On RTTS, Dehaze-RetinexGAN reaches NIMA 4.5313 (best) and BRISQUE 18.2351 vs RIDCP 17.4362 BRISQUE.
• On Fattal, NIMA improves by 3.77% and SSEQ by 13.00% over the best competing method per the authors.
• Uses about 10.00M parameters and 0.0093 s per image at 256x256, the fastest in the efficiency table.

Measured effects reported in the paper

Surprising findings

• The w/o Exposure Loss variant obtains a better BRISQUE (18.7454) than the w/o SSIM and w/o contrastive variants but the full model is still best overall.
• Pixel-level 'Ours' does not win every metric: on RTTS BRISQUE, RIDCP (17.4362) is better than Dehaze-RetinexGAN (18.2351).

Limitations

• Only non-reference quality metrics (BRISQUE, NIMA, SSEQ) are used; no full-reference evaluation on real data.
• Training pre-stage uses illumination-reflectance decomposition pseudo-labels from a traditional method, which may inject errors.
• Some metrics are not best (e.g., RIDCP better BRISQUE on RTTS and URHI).

Experimental setup

Datasets: URHI SOTS

Benchmarks: RTTS Fattal HSTS URHI test

Metrics: BRISQUE NIMA SSEQ running time parameter count

Baselines: DCP MSCNN DehazeNet AODNet MSBDN DAD FFANet PSD RefineDNet D4 Dehamer RIDCP

Builds on: Retinex theory atmospheric scattering model CycleGAN U-Net dark channel prior

Closest papers by meaning

• FD-GAN: Generative Adversarial Networks with Fusion-Discriminator for Single Image Dehazing (AAAI 2020 · similarity 0.91)
• Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing (CVPR 2025 · similarity 0.88)
• PromptHaze: Prompting Real-world Dehazing via Depth Anything Model (AAAI 2025 · similarity 0.88)
• Tokenize Image Patches: Global Context Fusion for Effective Haze Removal in Large Images (CVPR 2025 · similarity 0.86)
• Real-world Image Dehazing with Coherence-based Pseudo Labeling and Cooperative Unfolding Network (NeurIPS 2024 · similarity 0.86)
• IniRetinex: Rethinking Retinex-type Low-Light Image Enhancer via Initialization Perspective (AAAI 2025 · similarity 0.86)