🧵 Surface Defect Detection Demo — Textile Production

This repository contains a supervised deep-learning demo for pixel-level defect segmentation in industrial textile images.
A U-Net model, trained on manually labeled data, predicts surface anomalies. The demo includes a Flask web app where you can upload an image and instantly view the predicted defect mask side-by-side with the original.

---

🎯 Demo Highlights

Feature	Details
Architecture	U-Net (encoder–decoder with skip connections)
Learning type	Supervised segmentation — model trained on expert-annotated masks
Use case	Inline quality control / surface-defect detection in textile production
Web UI	Modern upload page, responsive layout, original & mask displayed together

---

Sample data
Ready-made test images reside in Test/ (masks in Test/Label/).
Upload any file from this folder through the web interface to see the model in action.

---

🚀 Quick Start

#  install requirements
pip install -r requirements.txt

#  run the web demo
python3 src/demo/backend.py

Open http://127.0.0.1:5000 and either upload your own 256 × 256 grayscale textile image or pick one from the Test/ folder.

Original	Predicted Mask

Example Output

🧠 Model Training Summary

1. Dataset DAGM 2007 (Class 1) resized to 256 × 256 grayscale
2. Labels Binary defect masks (1 = defect, 0 = background)
3. Training U-Net, 50 epochs, Adam (lr 1e-4), BCELoss
4. Validation 80 / 20 split; best weights saved as unet_dagm_class1.pth

Note: This is supervised ML — the network learns strictly from labeled examples.

---

⚙️ Adapting to Your Production Line

• Different materials? Fine-tune with your own labeled dataset.
• Other defect types? Retrain with additional mask channels.
• Edge deployment? Containerized builds available for on-prem devices.

For professional integration or custom training, contact us at info@sagel-ai.com

---