This paper proposes an AI vision-based defect detection system designed for manufacturing processes and presents its performance evaluation under realistic industrial conditions. The proposed system integrates image acquisition, automatic labeling, defect detection, and result analysis into a unified framework, and is implemented on an edge computing platform to enable real-time processing.
To evaluate the performance of the proposed system, a total dataset of 1,000 images was constructed, and each repeated experiment was conducted using 100 test samples under identical conditions. The evaluation metrics include defect recognition accuracy, defect type classification accuracy, object detection rate, and detection speed. Experimental results show that the proposed system achieves an average defect recognition accuracy of 94.6%, a defect type classification accuracy of 100%, and an object detection rate of 97.86%. In addition, the average defect detection time was measured as 0.073 seconds, demonstrating its suitability for real-time industrial applications.
However, limitations of vision-based inspection were observed when applied to diverse and complex industrial environments, particularly due to insufficient variability in training data. To address this issue, additional experiments were conducted using synthetic images generated from normal and defective samples collected from consortium partners. The results indicate that the use of synthetic data can enhance data diversity and improve model generalization performance.
The proposed system demonstrates high potential for practical deployment in manufacturing environments. Future work will focus on expanding the dataset, improving synthetic data generation techniques, and integrating MLOps-based continuous learning frameworks for adaptive process optimization.