Development of an Automated Computer Vision Model for Industrial Defect Detection

The client faces difficulties in efficiently and accurately detecting defects in manufactured metal components, such as inclusions, patches, and scratches. Manual visual inspection is time-consuming, inconsistent, and prone to human error, leading to increased defect rates, rework costs, and production delays. They require a scalable, automated solution to improve defect detection accuracy and speed while minimizing false positives that could disrupt production lines.

A mid-sized manufacturing company specializing in metal fabrication, seeking to implement machine learning-based quality inspection processes.

• Develop a computer vision model capable of detecting and classifying metal surface defects with an mAP of at least 85%.
• Achieve a recall and precision both above 80%, ensuring high detection sensitivity without excessive false positives.
• Enable rapid training, testing, and deployment of the model to facilitate quick iteration and validation.
• Provide an easy-to-use prediction endpoint accessible via API for integration into existing manufacturing workflows.
• Allow model deployment on edge devices for real-time inspection directly on the production line.

• Batch upload capability supporting image transformation (resizing to 256×256 pixels, RGB conversion)
• Annotation management for bounding boxes around defect types
• Model training interface with adjustable training budget (e.g., hours of training)
• Performance evaluation metrics including mAP, recall, and precision
• Prediction threshold adjustment to optimize detection sensitivity
• Model deployment with publishing and endpoint management for real-time inference
• Integration with external data sources and manufacturing control systems

• Manufacturing line data collection systems for automatic image capture
• Prediction API endpoints for real-time defect alerting
• Data storage solutions for image and annotation history

• System scalability to handle large volumes of images and predictions
• Inference latency suitable for real-time inspection (e.g., < 200ms per prediction)
• Security measures for protected data transfer and storage
• Model accuracy metrics with continuous monitoring to ensure performance stability

The project aims to significantly improve defect detection accuracy and throughput, reducing manual inspection time and human error. Achievement of at least 85% mAP, with recall and precision above 80%, will enable more consistent quality control, reduce rework and scrap costs, and increase overall production line efficiency. Deployment on edge devices will facilitate real-time defect identification, minimizing downtime and ensuring high-quality outputs aligned with industry standards.