Research map / 3D vision / Depth and stereo
Depth Information Assisted Collaborative Mutual Promotion Network for Single 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 DCMPNet, a dual-task network that jointly trains single-image dehazing and depth estimation, using depth-map differences between dehazed and GT images and dehazed-vs-GT residuals to mutually guide the two tasks.
Problem
Single image dehazing ignores the potential of depth information and downstream-task feedback to improve restoration of clear images from hazy inputs.
Approach
A difference-perception dual-task interaction: depth-map differences between the dehazed result and GT weight the dehazing loss, while the dehazed-vs-GT residual weights the depth estimation loss, so depth estimation and dehazing are optimized jointly and feed each other depth features.
Key results
- On SOTS-indoor the proposed method reaches 42.18 PSNR and 0.9967 SSIM, with best NIQE, PIQE and FADE among compared methods.
- On SOTS-outdoor it reaches 36.56 PSNR and 0.9931 SSIM, with best NIQE 2.9377, PIQE 7.6498 and FADE 0.7112.
- Ablation on SOTS-indoor: baseline 35.69 PSNR rises to 42.18 with all modules, including +4.72 dB from LEGM and +0.63 dB from DE-based dual-task learning.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LEGM (ablation) | Baseline dehazing network · SOTS-indoor | PSNR: 35.69 -> 40.41 | Helped |
| add MFM (ablation) | Baseline + LEGM · SOTS-indoor | PSNR: 40.41 -> 40.60 | Helped |
| add MSAAM (ablation) | Baseline + LEGM + MFM · SOTS-indoor | PSNR: 40.60 -> 40.75 | Helped |
| add depth estimation network (DE) (ablation) | Baseline + LEGM + MFM + MSAAM · SOTS-indoor | PSNR: 40.75 -> 41.55 | Helped |
| add difference perception (DP) (ablation) | Baseline + LEGM + MFM + MSAAM + DE · SOTS-indoor | PSNR: 41.55 -> 42.18 | Helped |
Surprising findings
- MB-TaylorFormer achieves higher PSNR on SOTS-outdoor (38.09) than the proposed method (36.56), and C2PNet/MB-TaylorFormer exceed it on indoor PSNR (42.56/42.63), so the method is not best on PSNR.
Limitations
- Gains on PSNR are not consistently best; comparable rather than superior on PSNR.
- Depth supervision relies on RA-Depth pseudo depth maps for ground truth depth rather than true depth labels.
- Evaluation limited to synthetic RESIDE benchmarks plus a small self-collected real set without quantitative real-world metrics.
Experimental setup
Datasets: RESIDE ITS RESIDE OTS SOTS-indoor SOTS-outdoor
Benchmarks: SOTS-indoor SOTS-outdoor real-world hazy images (self-collected)
Metrics: PSNR SSIM NIQE PIQE FADE
Baselines: AECR-Net PSD MAXIM-2S Dehamer D4 Dehazeformer C2PNet MB-TaylorFormer MITNet
Builds on: atmospheric scattering model RA-Depth AECR-Net VGG perceptual loss
Closest papers by meaning
- Depth-Centric Dehazing and Depth-Estimation from Real-World Hazy Driving Video (AAAI 2025 · similarity 0.89)
- PromptHaze: Prompting Real-world Dehazing via Depth Anything Model (AAAI 2025 · similarity 0.87)
- SelfPromer: Self-Prompt Dehazing Transformers with Depth-Consistency (AAAI 2024 · similarity 0.87)
- Adaptive Dynamic Dehazing via Instruction-Driven and Task-Feedback Closed-Loop Optimization for Diverse Downstream Task Adaptation (AAAI 2026 · similarity 0.86)
- Real-world Image Dehazing with Coherence-based Pseudo Labeling and Cooperative Unfolding Network (NeurIPS 2024 · similarity 0.86)
- Dehaze-RetinexGAN: Real-World Image Dehazing via Retinex-based Generative Adversarial Network (AAAI 2025 · similarity 0.85)
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.