IEEE Access (Jan 2024)
Enhancing EfficientNet-YOLOv4 for Integrated Circuit Detection on Printed Circuit Board (PCB)
Abstract
Ensuring the quality and functionality of printed circuit boards (PCBs) during manufacturing requires precise, automated visual inspection. Detecting integrated circuits (ICs) on PCBs poses a significant challenge due to diverse component sizes, types, and intricate board markings that complicate accurate object detection. This study addresses this challenge by proposing an enhanced EfficientNet-YOLOv4 algorithm tailored explicitly for the IC detection of PCBs. Numerous modifications are integrated into YOLOv4, with the replacement of its original backbone by a robust feature extraction network, EfficientNetv2-L, and meticulous hyperparameter tuning, including variations in loss functions, anchor size configurations, and other training techniques. The methodology further incorporates diverse data augmentation techniques to enrich the training dataset and enhance the model’s generalization ability. Extensive experiments conducted in this study showed the efficacy and robustness of the algorithm in handling complex PCB layouts and varying lighting conditions, outperforming existing PCB inspection models. The proposed method, EfficientNetv2-L-YOLOv4, achieved an impressive F1-score of 99.22 with an inference speed of 0.14 s per image. The proposed method also performed well compared to EfficientNet-B7-FasterRCNN and the original YOLOv4; it attains an F1-score of 98.96 and an inference speed of 0.10 s per image (with a batch size of 4). These results highlight the significance of effective feature extraction networks for object detection. Beyond addressing IC detection challenges, this algorithm advances the fields of computer vision and object detection. The implementation of EfficientNetv2-L-YOLOv4 in real manufacturing scenarios holds promise for automating component inspections and potentially eliminating the need for human intervention.
Keywords
