An Automated Optical Inspection (Aoi) System for Three-Dimensional (3d) Defects Detection on Glass Micro Optical Components (Gmoc)

As wavelength division multiplexing (WDM) being widely deployed, large number of glass micro optical components (GMOC) are utilized in optical transceivers. Visual inspection for these GMOCs is a necessary and critical manufacturing step for quality and reliability control. However, due to the facts that the glass components are usually transparent and defects are typically small and located randomly in 3D, manual inspection is often labor intensive and time consuming while the automated optical inspection (AOI) hasn’t been able to provide desirable accuracy and efficiency. In this paper, an AOI system incorporating 3D video acquisition and a novel machine-learning algorithm based on a two-stage neural network was developed successfully for 3D defects detection on GMOCs. It consists of a robotic arm for moving the parts in 3D, a camera with an illumination module for video acquisition of a part moving in 3D, and a video streaming processing unit empowered by a machine vision algorithm to detect the defects on GMOCs in real time on a production line. The robotic arm enables the fixed camera capture multi-perspective video of a test sample without having to refocus. The two-stage machine learning network is based on a modified YOLOv4 architecture with addition of color channel separation (CCS) convolution, an image quality evaluation (IQE) module, and frame fusion module to integrate the single frame detection results. It is capable of processing the multi-perspective video stream in a coarse-to-fine manner in real time for defects detection. Trained with 30 samples, the AOI system achieved very promising performances with a recall rate of 1, a detection accuracy of 97%, and an inspection time of 48s per part.