Efficient Printed Circuit Board Defect Segmentation Technology Based on YOLO Prompts and ICT-ViT

Table 1. Quantitative comparison of ViT models
View table
Table 1. Quantitative comparison of ViT models
ViT model mIoU mScore Inference time /ms Parameters /10⁶
ViT-h 0.973 0.966 16.2 637
ViT-l 0.955 0.962 14.7 308
ViT-b 0.936 0.892 12.0 90
TinyViT-21M 0.971 0.957 10.9 28
EfficientViT-M5 0.952 0.921 8.7 13
ICT-ViT 0.976 0.987 9.9 35

Table 2. Comparison of evaluation indexes of different mask modes

View table

Table 2. Comparison of evaluation indexes of different mask modes

Mask mode	Adjust parameter	Score						Inference time /ms
Mask mode	Adjust parameter	Missing_hole	Mouse_bite	Open_curcuit	Short	Spur	Spurious_copper	Inference time /ms
point	mask 1 is 20%	0.910	0.822	0.909	0.891	0.923	0.920	11.9
	mask 2 is 40%	0.925	0.855	0.898	0.858	0.891	0.884	10.3
	mask 3 is 60%	0.590	0.452	0.621	0.538	0.606	0.601	10.5
bbox	mask 1 is 20%	0.920	0.852	0.940	0.922	0.934	0.916	11.2
	mask 2 is 40%	0.875	0.882	0.887	0.887	0.936	0.902	10.7
	mask 3 is 60%	0.655	0.509	0.598	0.626	0.663	0.604	10.7

Table 3. Ablation experiments

View table

Table 3. Ablation experiments

Module	mIoU	mScore	Inference time /ms	Parameters /10⁶
None	0.920	0.832	11.9	32.22
YOLOv8-PCB	0.945	0.868	11.2	35.23
SSA	0.942	0.855	10.2	30.52
ICT-ViT	0.955	0.895	10.5	39.17
Adapters	0.941	0.848	10.7	32.52
YOLOv8-PCB+SSA	0.950	0.845	10.5	33.53
YOLOv8-PCB+SSA+ICT-ViT	0.965	0.870	9.8	40.48
YOLOv8-PCB+SSA+ICT-ViT+Adapters	0.976	0.889	9.7	40.78

Table 4. Comparison of defect segmentation of small targets of solid templates

View table

Table 4. Comparison of defect segmentation of small targets of solid templates

Defect class	MobileSAMv2			Proposed
Defect class	mIoU	mScore	Inference time /ms	mIoU	mScore	Inference time /ms
Missing_hole	0.872	0.845	10.9	0.900	0.801	9.8
Mouse_bite	0.943	0.999	13.2	0.910	0.977	12.7
Open_curcuit	0.850	0.814	15.9	0.941	0.946	11.7
Short	0.808	0.825	17.0	0.905	0.914	14.0
Spur	0.805	0.821	13.3	0.884	0.919	11.9
Spurious_copper	0.787	0.842	10.9	0.917	0.960	10.3

Table 5. Comparison of performance of different models on solid templates
View table
Table 5. Comparison of performance of different models on solid templates
Model mIoU mScore Inference time /ms Model size /MB
FastSAM-s 0.863 0.721 12.6 22.7
EfficientSAM 0.881 0.835 11.2 39.0
MobileSAMv2 0.844 0.858 13.5 38.8
Proposed 0.910 0.920 11.7 50.2

Table 6. Comparison of the influence of different production process interference on model performance

View table

Table 6. Comparison of the influence of different production process interference on model performance

Disturbance type	MobileSAMv2			Proposed
Disturbance type	mIoU	mScore	Inference time /ms	mIoU	mScore	Inference time /ms
Normal	0.920	0.832	11.9	0.976	0.889	9.7
Dent	0.621	0.773	10.6	0.624	0.765	9.3
Scratch	0.612	0.652	10.8	0.783	0.635	9.2
Spot	0.803	0.852	10.0	0.891	0.935	9.0

Tools

Get Citation

Copy Citation Text

Tieqiang Sun, Hongjian Yu, Can Zhang, Yidong Yuan, Aoran Sun. Efficient Printed Circuit Board Defect Segmentation Technology Based on YOLO Prompts and ICT-ViT[J]. Laser & Optoelectronics Progress, 2025, 62(12): 1215008

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites