🔧 Welding Quality Detection Challenge

Welcome to the official repository for our solution to the Industrial and Trustworthy AI Challenge: Welding Quality Detection, hosted by Confiance.ai, Renault Group, and IRT SystemX.

---

---

🚀 Objective

Develop a trustworthy AI component that automatically classifies welding seams from automotive production line images into:

• ✅ OK: Weld is normal.
• ❌ KO: Weld has defects.
• ❓ UNKNOWN: AI is uncertain.

The model assists human operators by reducing manual inspections and improving safety.

---

📦 Challenge Highlights

• Dataset: 22,753 weld images (C20, C33, C102)
• Imbalance: ~98% OK vs ~2% KO
• Variability: Lighting, blur, angle, position
• Outputs:
  • Class: OK, KO, UNKNOWN
  • Probability per class (optional)
  • OOD score for distribution awareness

---

🧠 Participant's Approach (Work-in-Progress)

This is an evolving solution. Core directions include:

• 🔍 CNN-based classifier (ResNet/EfficientNet)
• 🔁 Transfer learning for limited KO data
• 🎨 Augmentation: rotation, blur, contrast, motion
• 🎯 Uncertainty estimation (e.g., MC-Dropout)
• ⚠️ OOD detection using softmax entropy + VAE
• 📏 Feature extraction from welds (length, width, area)
• 🧠 Explainability using GradCAM

---