Fig. 1. The architecture model of YOLOv8

Fig. 2. Small object detection head fused with shallow scale

Fig. 3. Conv downsampling and ADown downsampling structures

Fig. 4. Sampling based on dynamic upsampling

Fig. 5. Sampling point generator in Dysample

Fig. 6. The network architecture model of YOLOv8-PCB

Fig. 7. Types of PCB defects

Fig. 8. Loss function curve comparison

Fig. 9. Comparison of model recall rate and mAP@0.5 value curve before and after improvement

Fig. 10. The confusion matrix for the six types of defects. (a) Confusion matrix of YOLOv8; (b) confusion matrix of YOLOv8-PCB

Fig. 11. YOLOv8 defect detection results

Fig. 12. YOLOv8-PCB defect detection results