Research map / Image generation / Image restoration
Prior-guided Hierarchical Harmonization Network for Efficient 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 PGH²Net, a pure-convolutional UNet-like image dehazing network that injects bright/dark channel priors and histogram-equalization guidance via prior aggregation, harmonization, sandwich and HEGM modules.
Problem
Image dehazing with vision Transformers is costly (quadratic complexity) and existing methods lack dehazing priors, limiting efficiency and real-world applicability.
Approach
Injects bright/dark channel priors (B/DCP) computed on deep feature maps for spatial guidance and uses histogram-equalization similarity as 1D channel-wise guidance in the bottleneck, with a lightweight pure-convolution design.
Key results
- Achieves 41.70 dB PSNR and 0.996 SSIM on SOTS-Indoor with 1.76M parameters and 16.05G FLOPs.
- Reports 37.52 dB PSNR on SOTS-Outdoor and 17.02 dB on Dense-Haze, outperforming listed baselines.
- Ablation on SOTS-Indoor: sandwich module gives +1.40 dB and HEGM +0.63 dB over the baseline (34.69 dB).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the gating branch to the baseline configuration (ablation) | PGH²Net encoder-decoder modules · SOTS-Indoor | PSNR: 35.64 -> 36.27 | Helped |
| Add the Prior Aggregation (PA) module (ablation) | PGH²Net encoder-decoder modules · SOTS-Indoor | PSNR: 36.27 -> 37.14 | Helped |
| Add the Spatial Harmonization (SH) module (ablation) | PGH²Net encoder-decoder modules · SOTS-Indoor | PSNR: 37.14 -> 37.48 | Helped |
| Add the Channel Harmonization (CH) module (ablation) | PGH²Net encoder-decoder modules · SOTS-Indoor | PSNR: 37.48 -> 38.00 | Helped |
| Add the Sandwich module to the bottleneck without HEGM (ablation) | PGH²Net bottleneck · SOTS-Indoor | PSNR: 34.69 -> 36.09 | Helped |
| Add HEGM to the bottleneck without the Sandwich module (ablation) | PGH²Net bottleneck · SOTS-Indoor | PSNR: 34.69 -> 35.32 | Helped |
| Use both the Sandwich module and HEGM instead of neither (ablation) | PGH²Net bottleneck · SOTS-Indoor | PSNR: 34.69 -> 38.00 | Helped |
| Use PGH²Net rather than TaylorFormer | Image dehazing models · SOTS-Outdoor | PSNR: 37.42 -> 37.52 | Helped |
| Use PGH²Net rather than LH-Net | Image dehazing models · Dense-Haze | PSNR: 18.87 -> 17.02 | Hurt |
| Use PGH²Net rather than TaylorFormer | Image dehazing models · O-HAZE | PSNR: 25.05 -> 25.47 | Helped |
Limitations
- Ablations are reported only on SOTS-Indoor.
- Hardware and training time are not reported.
Experimental setup
Datasets: RESIDE-Indoor RESIDE-Outdoor SOTS-Indoor SOTS-Outdoor Dense-Haze O-HAZE
Benchmarks: SOTS-Indoor SOTS-Outdoor Dense-Haze O-HAZE
Metrics: PSNR SSIM Params (M) FLOPs (G)
Baselines: DehazeNet AOD-Net GridDehaze MSBDN FFA-Net AECR-Net DeHamer PMNet MAXIM-2S DehazeFormer TaylorFormer LH-Net
Builds on: dark channel prior TaylorFormer ConvNeXt
Closest papers by meaning
- Exploiting Diffusion Prior for Real-World Image Dehazing with Unpaired Training (AAAI 2025 · similarity 0.87)
- RIDCP: Revitalizing Real Image Dehazing via High-Quality Codebook Priors (CVPR 2023 · similarity 0.87)
- PromptHaze: Prompting Real-world Dehazing via Depth Anything Model (AAAI 2025 · similarity 0.86)
- Data-independent Module-aware Pruning for Hierarchical Vision Transformers (ICLR 2024 · similarity 0.85)
- Dehaze-RetinexGAN: Real-World Image Dehazing via Retinex-based Generative Adversarial Network (AAAI 2025 · similarity 0.85)
- Depth Information Assisted Collaborative Mutual Promotion Network for Single Image Dehazing (CVPR 2024 · 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.