PCB Defect Detection Method Based on G-YOLOv8

In the electronics industry, PCB is crucial for ensuring the product quality. However, these defects of PCB generally have the characteristics of complex backgrounds, small sizes and similar features, resulting in the issues of high computational complexity or low detection precision of PCB defect detection algorithms based on deep learning. To resolve the aforementioned concerns, this paper proposed a lightweight PCB defect detection model named G-YOLOv8 which was based on the YOLOv8. The G-YOLOv8 can effectively reduce the computational complexity by integrating the Ghost Conv into the Backbone and Neck networks. Furthermore, the G-C2f module was introduced into Neck network to strengthen the capability of G-YOLOv8 to perceive the feature information at different scales and semantic levels, thus improving the detection performance of the model for the PCB defect. Meanwhile, the performance of G-YOLOv8 model was evaluated on the public data set. The experimental results show that the parameter number of G-YOLOv8 is 2.6 million, which is decreased by 89.7% and 13.3% compared with SSD and YOLOv8, respectively. The computational load of G-YOLOv8 is 7.5 GFLOPs, which is decreased by 93.5% and 7.4% compared with SSD and YOLOv8, respectively. The [email protected] of G-YOLOv8 is 94.4%, which is increased by 4.8% and 1.3% compared with SSD and YOLOv8, respectively. Hence, in the task of PCB defect detection, the G-YOLOv8 model can reduce the computational complexity while preserving a high detection precision, which can meet the requirements of PCB defect detection in industrial scenarios.