A real-time PCB defect detection model based on enhanced semantic information fusion

Aiming at the problem of poor balance between real-time and performance of existing PCB board defect detection algorithms, this paper proposes a fast PCB board detection model based on enhanced semantic information fusion, i.e., the Ghost-YOLOv8 (G-YOLOv8) model. GhostConv is used for feature extraction in the backbone network part, which reduces the complexity of network operation; the SPPFCSPCS structure is proposed and applied to the deepest layer of the backbone network to strengthen the model’s fusion ability for deep multi-scale semantic information. In addition, after the feature fusion large-scale detection head, the low parameter number \(\textrm{A}^2\) ${\text{A}}^2$ attention mechanism is added to improve the model’s attention to high-dimensional valid semantic information and optimize the detection results of the network; the Wise-IoU loss function is used for iteration in the training process to strengthen the model’s fitting ability. The experimental results show that the G-YOLOv8 model can detect up to 125FPS on the Peking University open-source PCB board defect detection dataset and Deep PCB dataset, which is an improvement of 1.5FPS compared to the pre-improvement period; and the mAP0.5 on the deep PCB dataset improves by 0.8%. The performance and detection speed balance of the network has been improved.