Computer Vision-Based In-Line Inspection in Intelligent Quality Control Systems: Methods, Applications, and Engineering Challenges

Abstract

On-line visual inspection-that is, automatically taking images and analyzing them during manufacturing, rather than after manufacturing-is very important for intelligent quality control of modern factories. It replaces slow and subjective manual inspection with rapid and objective detection. This paper reviews the on-line industrial inspection method based on computer vision and deep learning, using the theory of network physical system (CPS), system engineering and information theory. Inspection chain is regarded as a channel to obtain information, and its capacity must be large enough to track production and find defects. It lays a theoretical foundation for intelligent quality control system with vision, which is regarded as closed-loop CPS, in which the fault signal from the image always adjusts the machine settings, not just that the product is defective. We analyzed four changes: from sample inspection to continuous whole surface inspection, from once discarding to predicting defects before they arrive, from manual judgment to closed-loop automatic quality control, and from single inspection station to integrated quality -CPS ecosystem. The applications in three main fields-surface defect detection of metal materials, unattended anomaly detection when defects are rare, and real-time quality control of edge equipment-are illustrated by the reference data set (MVTec AD), the latest architecture and factory use examples. There are still thorny problems such as class imbalance, the cost of tagging data, the difficulty of generalization from one domain to another, and the choice between speed and accuracy on edge devices. A new framework concept puts forward that the uncertainty of detection is regarded as an important signal to improve the production process, and the quality control is changed from simple "good or bad" to a means to continuously improve manufacturing.