Defect Detection of Printed Matter Based on Improved YOLOv5l

Haiwen Liu; Yuanlin Zheng; Chongjun Zhong; Kaiyang Liao; Bangyong Sun; Hanxiang Zhao; Jie Lin; Haoqiang Wang; Shanxiang Han; Bo Xie

doi:10.3788/LOP231826

Laser & Optoelectronics Progress, Volume. 61, Issue 10, 1012002(2024)

Defect Detection of Printed Matter Based on Improved YOLOv5l

Haiwen Liu^1、*, Yuanlin Zheng^1、**, Chongjun Zhong², Kaiyang Liao¹, Bangyong Sun¹, Hanxiang Zhao¹, Jie Lin¹, Haoqiang Wang¹, Shanxiang Han¹, and Bo Xie²

Author Affiliations

¹College of Faculty of Printing, Packaging Engineering and Digital Media Technology, Xi'an University of Technology, Xi'an 710054, Shaanxi , China

²Weinan Daily Printing Factory, Weinan 714099, Shaanxi , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(11)

Fig. 1. Architecture of YOLOv5l network

Download full size

Fig. 2. Architecture of improved YOLOv5l network

Download full size

Fig. 3. Omni-dimensional dynamic convolution

Download full size

Fig. 4. Structure diagram of C3Ghost module

Download full size

Fig. 5. Types of printing defects. (a) Satellite droplet; (b) spots; (c) missing print; (d) crack

Download full size

Fig. 6. Loss variation curve

Download full size

Fig. 7. Comparison of printing defect detection effects between YOLOv5l and improved YOLOv5l. (a) Detection results of YOLOv5l; (b) detection results of improved YOLOv5l

Download full size

Table 1. Experimental environment configuration

View table

Table 1. Experimental environment configuration

Category	Parameter	Configuration
Hardware	Video storage	11 GB
	Memory	16 GB
	GPU	NVIDIA GeForc RTX 2080 Ti
	CPU	Intel（R） Core（TM） i5-4590 CPU@3.30 GHz
Software	Operating system	Unbantu 22.04
	Programming language	Python 3.7
	Graphics card acceleration	CUDA 10.1
	Deep learning framework	PyTorch 1.10.1

Table 2. Number of various defects
View table
Table 2. Number of various defects
Types of printing defects Number of defects
Satellite droplet 976
Spots 985
Missing print 1041
Crack 998

Table 3. Experimental results of different improvement modules
View table
Table 3. Experimental results of different improvement modules
Model Four detection scales ODConv C3Ghost mAP /% Speed /（frame·s^-1）
YOLOv5l — — — 94.4 48.1
Improvement 1 √ — — 96.5 40.8
Improvement 2 √ √ — 97.7 39.1
Improvement 3 √ √ √ 97.3 44.1

Table 4. Performance comparison of different target detection algorithms

View table

Table 4. Performance comparison of different target detection algorithms

Model	AP（R_IOU=0.5）				mAP
Model	Satellite droplet	Spots	Missing print	Crack	mAP
EfficientDet^［18］	89.1	88.7	84.3	79.8	85.5
Faster R-CNN^［18］	90.3	93.8	87.1	84.2	88.8
SSD^［18］	82.5	90.4	54.0	63.7	72.7
YOLOv3^［18］	87.2	90.9	82.3	83.2	85.9
YOLOv4^［18］	90.5	92.1	87.5	85.7	88.9
Siamese-YOLOv4^［18］	98.2	96.5	92.0	91.7	94.6
YOLOv5l	95.9	99.5	94.2	87.9	94.4
Proposed model	99.4	99.5	95.6	94.7	97.3

Tools

Get Citation

Copy Citation Text

Haiwen Liu, Yuanlin Zheng, Chongjun Zhong, Kaiyang Liao, Bangyong Sun, Hanxiang Zhao, Jie Lin, Haoqiang Wang, Shanxiang Han, Bo Xie. Defect Detection of Printed Matter Based on Improved YOLOv5l[J]. Laser & Optoelectronics Progress, 2024, 61(10): 1012002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites