Comparison of Deep Learning Models for Singlsyn Defect Detection and Classification

The use of computers for inspecting Singlsyn parts before shipment has resulted in occasional errors when distinguishing between usable (OK) and non-usable (NG) parts. Initially, the company's classification system exhibited inaccuracies, sometimes misclassifying OK parts as NG, achieving an accuracy rate of 90%, necessitating manual sorting of the misclassified parts, which was time-consuming. To enhance the accuracy and efficiency in differentiating between OK and NG parts-an inherently challenging task-deep learning techniques were introduced, particularly object detection methods. The target accuracy for this research was set at 97%. A total of eight models were tested, with various parameters adjusted throughout the process. The SSD-Mobilenet-V2-FPNLite-320 model delivered the best results, improving the system's accuracy in detecting OK and NG parts from 90% to 99.58%.