Monodara/steel-defect-detection

   

An AI-driven prototype designed for automated surface quality assurance in the steel industry. This project utilizes a fine-tuned YOLOv11 model to identify and localize typical surface defects with high precision.

🌟 Key Features

• —Real-time Inference: Upload images for near-instantaneous defect detection.
• —Multi-class Recognition: Specifically trained to detect defects like Crazing, Inclusion, Patches, Pitted Surfaces, Rolled-in Scale, and Scratches.
• —Adjustable Sensitivity: Sidebar controls for Confidence Threshold to adapt to various industrial tolerance levels.
• —Robust Deployment: Optimized Docker environment with refined OpenCV dependencies for cloud-based CPU inference.

📊 Methodology & Performance

• —Base Architecture: YOLOv11s (Small variant for optimal speed-accuracy trade-off).
• —Training Strategy: Iterative fine-tuning (v1 to v5) with focus on small-feature recall and textured background suppression.
• —Input Resolution: Optimized at 640px to balance CPU latency (~1s) and defect detection sensitivity.
• —Inference Engine: Seamless integration of PyTorch and Ultralytics.

🛠️ Tech Stack

• —Algorithm: Ultralytics / PyTorch
• —Interface: Streamlit
• —Containerization: Docker
• —Processing: OpenCV-headless

📖 How to Use

1. 1.Adjust the Confidence Threshold in the sidebar (Recommended: 0.35 - 0.50).
2. 2.Upload a steel surface image (JPG/PNG).
3. 3.The system will render a bounding box overlay and provide a Defect Inventory report below the result.