PHFD-Net: A Locally Re-parameterized Hierarchical Spatial-Frequency Network for Single Image Dehazing
Project page for PHFD-Net. The paper is currently under review at IEEE TCSVT.
This repository shares paper-level information — method overview, results, figures, and evaluation-protocol notes. Source code, pretrained models, and evaluation scripts will be released upon acceptance (see docs/release_plan.md).
PHFD-Net is a hierarchical single-image dehazing network that combines local detail-enhanced restoration, deep-state residual routing, and a spatial-frequency bottleneck. Local branches are re-parameterizable, so the deployed model runs as a compact single-branch graph (4.74 M parameters) without changing its outputs.
- local detail-enhanced restoration;
- deep-state history residual routing (PreRoute-SF);
- spatial-frequency bottleneck refinement;
- local-branch re-parameterization for efficient deployment.
- Benchmark comparison: results/quantitative_results.md
- Ablation study: results/ablation_results.md
- Complexity and speed: results/complexity_profile.md
Benchmarks: SOTS-Indoor, Haze4K, and SOTS-Outdoor. Fine-tuned results are reported separately from matched/scratch comparisons. The SOTS-Outdoor row uses the Pillow 8.3.2 JPEG decoding protocol.
See docs/release_plan.md. Code and pretrained models will be released after the paper is accepted.
The bibliographic information will be updated once the paper is published. See CITATION.cff.
Released under the MIT License.