FACS-Net: Frequency-Aware Crack Segmentation Network for Thin Cracks (Checkpoint, and Results)

Frequency-Aware Crack Segmentation Network (FACS-Net) for Thin Cracks via Topology Preservation
FACS-Net is a frequency-aware neural architecture designed to enhance the segmentation of ultra-thin cracks (≤2px) in civil infrastructure images. It combines a hybrid CNN-Transformer encoder with a frequency-modulated decoder and introduces a topology-aware loss function (CT-Loss) to enforce structural continuity.

Contents

🔧 Pretrained Models- `CV12K.ckpt` – trained on CrackVision12K
- `Omni.ckpt` – trained on OmniCrack30K

🖼️ Segmentation Outputs- `CrackVision12K.zip` – result images from CrackVision12K test set
- `OmniCrack30K.zip` – result images from OmniCrack30K benchmark

📊 Evaluation- Includes metrics: IoU, CL-IoU, CTS
- Use with the official codebase for inference and analysis

> 🔗 Project GitHub Repository:https://github.com/SH-Joo/FACS

---

License:
All contents are released under the MIT License. Free to use and modify with proper attribution.

Note: The associated paper is currently under review. Please cite the SSRN preprint when available.