Research map / Label-efficient and robust learning / Self-supervised and contrastive
ODCR: Orthogonal Decoupling Contrastive Regularization for Unpaired Image Dehazing
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 ODCR, an unpaired image dehazing method that decouples generator features into haze-related and haze-unrelated components using Stiefel-manifold orthogonal MLPs, a self-supervised depth-wise feature classifier, and a weighted PatchNCE loss.
Problem
Unpaired image dehazing suffers from the conflict between CUT-style mutual-information maximization (pulling output patches toward hazy input) and actually removing haze, since haze-related and haze-unrelated features are entangled.
Approach
Orthogonally decouple patch features with MLPs optimized on the Stiefel manifold, use a depth-wise classifier to weight channels as haze-related or unrelated in a self-supervised way, and apply a weighted PatchNCE that pulls haze-related components toward clear patches and haze-unrelated components toward the hazy input.
Key results
- ODCR reaches 26.32 dB PSNR / 0.945 SSIM on SOTS-indoor, the best unpaired result and second overall behind supervised GDN.
- On SOTS-outdoor and NH-HAZE 2 (not trained on), ODCR reports 26.16 dB and 17.56 dB PSNR, outperforming all compared paired and unpaired methods including GDN.
- Ablation: replacing PatchNCE with WPNCE raises PSNR from 24.30 to 26.32 dB; Stiefel-manifold O-MLP beats an orthogonal loss (24.96 dB) and no decoupling (22.30 dB).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ODCR instead of GDN | Unpaired image dehazing · SOTS-indoor | PSNR (dB): 32.16 -> 26.32 | Hurt |
| Use ODCR instead of GDN | Unpaired image dehazing · SOTS-indoor | SSIM: 0.983 -> 0.945 | Hurt |
| Use ODCR instead of D4 | Unpaired image dehazing · SOTS-outdoor | PSNR (dB): 25.83 -> 26.16 | Helped |
| Use ODCR instead of D4 | Unpaired image dehazing · SOTS-outdoor | SSIM: 0.956 -> 0.960 | Helped |
| Use ODCR instead of CUT | Unpaired image dehazing · NH-HAZE 2 | PSNR (dB): 15.92 -> 17.56 | Helped |
| Use ODCR instead of CUT | Unpaired image dehazing · NH-HAZE 2 | SSIM: 0.758 -> 0.766 | Helped |
| Add orthogonal decoupling optimized on the Stiefel manifold (ablation) | ODCR · SOTS-indoor | PSNR (dB): 22.30 -> 26.32 | Helped |
| Add orthogonal decoupling optimized on the Stiefel manifold (ablation) | ODCR · SOTS-indoor | SSIM: 0.887 -> 0.945 | Helped |
| Optimize the orthogonal constraint on the Stiefel manifold rather than using an orthogonal loss (ablation) | ODCR · SOTS-indoor | PSNR (dB): 24.96 -> 26.32 | Helped |
| Replace DWFC with fixed feature assignment ratios (ablation) | ODCR · SOTS-indoor | PSNR (dB): 26.32 vs 13.10; 13.93; 15.33; 14.30; 18.28 | Hurt |
| Replace DWFC with fixed feature assignment ratios (ablation) | ODCR · SOTS-indoor | SSIM: 0.945 vs 0.599; 0.574; 0.621; 0.660; 0.726 | Hurt |
| Replace WPNCE with PatchNCE (ablation) | ODCR · SOTS-indoor | PSNR (dB): 24.30 -> 26.32 | Hurt |
Surprising findings
- Fixed ratio assignment of haze-related vs unrelated channels performed far worse (13.10 to 18.28 dB) than the learned DWFC assignment.
Limitations
- Evaluated only on a few dehazing datasets with training on ITS alone.
- Decoupling relies on the assumption that features can be separated into haze-related and unrelated components, which the paper admits is hard due to deep coupling in the ASM.
Experimental setup
Datasets: RESIDE ITS
Benchmarks: SOTS-indoor SOTS-outdoor NH-HAZE 2 Fattal's dataset
Metrics: PSNR SSIM parameters (M) latency (ms)
Baselines: DehazeNet AOD-Net MSCNN GDN DCP CycleGAN CycleDehaze YOLY USID-Net RefineDNet D4 CUT
Builds on: CUT CycleGAN AECR D4
Closest papers by meaning
- Contrastive Learning for Low-Light Raw Denoising (Student Abstract) (AAAI 2024 · similarity 0.84)
- Exploiting Diffusion Prior for Real-World Image Dehazing with Unpaired Training (AAAI 2025 · similarity 0.84)
- Dehaze-RetinexGAN: Real-World Image Dehazing via Retinex-based Generative Adversarial Network (AAAI 2025 · similarity 0.83)
- Iterative Predictor-Critic Code Decoding for Real-World Image Dehazing (CVPR 2025 · similarity 0.83)
- Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing (CVPR 2025 · similarity 0.83)
- Prior-guided Hierarchical Harmonization Network for Efficient Image Dehazing (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.