Defect Detection of Chip on Carrier Based on Improved YOLOV3

Tianyu Zhou; Qibing Zhu; Min Huang; Guiliang Cai; Xiaoxiang Xu

doi:10.3788/LOP202158.1210002

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 12 >Page 1210002 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 12, 1210002 (2021)

Defect Detection of Chip on Carrier Based on Improved YOLOV3

Tianyu Zhou¹, Qibing Zhu^1、*, Min Huang¹, Guiliang Cai¹, and Xiaoxiang Xu²

Author Affiliations

¹Key Laboratory of Advanced Process Control for Light Industry, Ministry of Education, Jiangnan University, Wuxi, Jiangsu 214122, China

²Wuxi CK Electric Control Equipment Co., Ltd., Wuxi, Jiangsu 214400, China

show less

DOI: 10.3788/LOP202158.1210002 Cite this Article Set citation alerts

Tianyu Zhou, Qibing Zhu, Min Huang, Guiliang Cai, Xiaoxiang Xu. Defect Detection of Chip on Carrier Based on Improved YOLOV3[J]. Laser & Optoelectronics Progress, 2021, 58(12): 1210002 Copy Citation Text

show less

Fig. 1. Intact COC

Download full size

Fig. 2. COC defect

Download full size

Fig. 3. Structure of YOLOV3

Download full size

Fig. 4. Multi-scale feature fusion of YOLOV3

Download full size

Fig. 5. Improved multi-scale feature fusion

Download full size

Fig. 6. Extraction of the area to be detected on COC. (a) Original image; (b) edge detection; (c) corner detection; (d) area to be detected on COC

Download full size

Fig. 7. Loss value graph

Download full size

Fig. 8. COC defect detection results based on YOLOV3-COC. (a) Original image 1; (b) detection result 1; (c) original image 1; (d) detection result 2

Download full size

Res	Type	Filters	Size(stride)	Output
	Conv	32	3×3	416×416×32
	Conv	64	3×3(2)	208×208×64
1×	Conv	32	1×1
	Conv	64	3×3
	Residual			208×208×64
	Conv	128	3×3(2)	104×104×128
2×	Conv	64	1×1
	Conv	128	3×3
	Residual			104×104×128
	Conv	256	3×3(2)	52×52×256
8×	Conv	128	1×1
	Conv	256	3×3
	Residual			52×52×256
	Conv	512	3×3(2)	26×26×512
8×	Conv	256	1×1
	Conv	512	3×3
	Residual			26×26×512
	Conv	1024	3×3(2)	13×13×1024
4×	Conv	512	1×1
	Conv	1024	3×3
	Residual			13×13×1024

Table 1. Darknet-53

Res	Type	Filters	Size(stride)	Output
	Conv	32	3×3	576×576×32
	Conv	64	3×3(2)	288×288×64
1×	Conv	32	1×1
	Conv	64	3×3
	Residual			288×288×64
	Conv	128	3×3(2)	144×144×128
2×	Conv	64	1×1
	Conv	128	3×3
	Residual			144×144×128
	Conv	256	3×3(2)	72×72×256
8×	Conv	128	1×1
	Conv	256	3×3
	Residual			72×72×256
	Conv	512	3×3(2)	36×36×512
4×	Conv	256	1×1
	Conv	512	3×3
	Residual			36×36×512
	Conv	1024	3×3(2)	18×18×1024
4×	Conv	512	1×1
	Conv	1024	3×3
	Residual			18×18×1024
	Conv	2048	3×3(2)	9×9×2048
2×	Conv	1024	1×1
	Conv	2048	3×3
	Residual			9×9×2048

Table 2. Improved feature extraction network Darknet-49

Feature map size	Size	Number
9×9	$(40 \times 230)$ ; $(58 \times 343)$ ; $(178 \times 156)$	9×9×3
18×18	$(108 \times 38)$ ; $(45 \times 136)$ ; $(18 \times 118)$	18×18×3
36×36	$(57 \times 77)$ ; $(90 \times 18)$ ; $(62 \times 45)$	36×36×3
72×72	$(6 \times 8)$ ; $(13 \times 10)$ ; $(16 \times 30)$	72×72×3

Table 3. A priori box

Defect	Total	Correct detection	Missed detection	Accuracy	Missed detection rate
Waveguide stain	69	68	1	0.986	0.014
Collapse	98	95	3	0.969	0.031
Positioning column damage	27	26	1	0.963	0.037
Total	194	189	5	0.974	0.026

Table 4. Defect detection statistics

Model	Backbone	Size	Accuracy /%	Time /s
Faster R-CNN	VGG-16	600×600	82.6	0.94
SSD	VGG-16	512×512	85.4	0.63
YOLO	Google-Net	448×448	74.8	0.65
YOLOV2	Darknet-19	416×416	76.3	0.68
YOLOV3	Darknet-53	416×416	92.5	0.76
YOLOV3-COC	Darknet-49	576×576	97.4	0.72

Table 5. COC defect detection results of different algorithms

Tianyu Zhou, Qibing Zhu, Min Huang, Guiliang Cai, Xiaoxiang Xu. Defect Detection of Chip on Carrier Based on Improved YOLOV3[J]. Laser & Optoelectronics Progress, 2021, 58(12): 1210002

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information