Metric Learning-Based Incorrect Component Detection Framework

Artificial intelligence is applied in surface mount technology (SMT) to reduce the false alarm rate of automatic optical inspection systems. Among SMT defects, incorrect component placement-where a component is positioned incorrectly-is one of the most critical and damaging issues. Many factories rely on optical character recognition (OCR) to detect incorrect components; however, OCR struggles with blurry, obstructed, or dirt-covered characters. To address these challenges, this paper proposes a Metric Learning-based Incorrect Component Detection (MICD) framework, designed to handle new component and text detection tasks in low-quality images. The MICD framework is evaluated against the OCR-based method, demonstrating a leakage rate below 7% and a false alarm rate of only 10% of that of OCR. Overall, the proposed framework effectively detects incorrect components across various production lines while significantly reducing false alarm rates.