Deep Learning-Based Classification Approach for Wire Bonding Defects Inspection

The wire bonding process is one of the most vital processes in semiconductor manufacturing. Therefore, defect detection is needed to ensure the quality of the produced integrated circuits (ICs), in which poor quality wire bonding can prevent it from functioning effectively. An automatic optical inspection (AOI) system is commonly used for defect inspection in fabrication mode. However, the AOI system suffers from a lot of challenges that require human assistance in the event of uncertain defect classification. In consequence, manual inspection leads to low productivity and can be influenced by human errors. Therefore, it is necessary to integrate the AOI with artificial intelligence (AI) technology to replace human assist for productivity and quality improvement. Hence, the main focus of this paper is to find the best deep learning model suitable to be used for wire bonding defect classification. Various deep learning models have been tested using the original images collected from semiconductor fabrication. From the experimental results, EfficientNetB0 V2 is selected as the best model to be used for defect inspection by considering the accuracy and processing speed with the best results of 98% and 0.045 seconds, respectively. Moreover, the model also retains its lightweight nature with model size of 29 MB which is tolerable to be deployed in the real-world application.