A Small-sample Defect Detection Method for Screen-printed Characters with Adaptive Target Region Extraction

doi:10.3969/j.issn.1004-132X.2025.12.020

Abstract

Abstract:

The defect detection of screen-printed characters on electronic devices faced problems such as poor segmentation accuracy caused by reflective materials， difficulty in adaptively locating and extracting target regions due to uncertain image poses， and low recognition accuracy caused by small defect detection targets and insufficient samples， an adaptive target region extraction method for few-shot screen-printed character defect detection was proposed. Based on OTSU method an adaptive dual-threshold segmentation algorithm was designed to reduce information loss in bright target regions on reflective material surfaces， achieving accurate segmentation of images with uneven illumination. An adaptive target region localization， extraction， and angle correction algorithm was proposed to solve the problems of precisely locating and extracting adaptive target regions despite variations in character sizes and poses. A method involving preliminary defect target recognition and secondary fine recognition for low-confidence small target characters was studied， achieving accurate recognition for small-target， few-sample defect detection targets. Experimental results demonstrate that： the dual-threshold segmentation algorithm achieves accurate segmentation of character images under uneven illumination； the accuracy of adaptive target region localization and angle correction reaches 99.5%； the character classification recognition rate of the lightweight deep learning model reaches 99.1%； the character defect detection accuracy reaches 98.6%； and the detection speed is as 0.083 seconds per image. These results meet the requirements for both of precision and speed in industrial online detections.

Key words: defect detection, screen printing, non-uniform lighting, pose variation, small sample recognition

摘要：

针对因电子器件丝印字符材质反光引起分割精度差、检测图像位姿不确定难以自适应定位提取目标区域、缺陷检测目标小及样本少导致识别准确率低的问题，提出自适应目标区域提取的小样本丝印字符缺陷检测方法。设计基于大津阈值法的自适应二次阈值分割算法，减少反光材质表面光亮目标区域信息损失，实现光照不均匀图像的准确分割；提出自适应目标区域定位提取与角度矫正算法，解决因字符大小及位姿变化难以实现自适应目标区域的精确定位与提取问题；研究了缺陷检测目标初识别、置信度低小目标字符二次精识别方法，实现了小目标、少样本缺陷检测目标的准确识别。实验结果表明：二次阈值分割算法能够实现光照不均字符图像的准确分割，自适应目标区域定位与角度矫正准确率达99.5%，轻量化深度学习模型的字符分类识别率达99.1%，字符缺陷检测准确率达98.6%，检测速度为0.083 s/张，满足工业生产中在线检测精度与速度的需求。

关键词: 缺陷检测, 丝网印刷, 光照不均匀, 位姿不确定, 小样本识别

CLC Number:

TP277

Xinyu HU, Junwei ZHANG, Jia AI, Yunling LI, Shuang YAN. A Small-sample Defect Detection Method for Screen-printed Characters with Adaptive Target Region Extraction[J]. China Mechanical Engineering, 2025, 36(12): 2968-2977.

胡新宇, 张骏巍, 艾佳, 李云翎, 严爽. 一种自适应目标区域提取的小样本丝印字符缺陷检测方法[J]. 中国机械工程, 2025, 36(12): 2968-2977.

Figures/Tables 25

Fig.1 Visual identity system

Fig.2 Algorithm flowchart

Fig.3 The physical platform of the system

Fig.4 Partial defect diagram

Fig.5 The physical platform of the system

Fig.6 The binarization of the improved algorithms

Fig.7 Binary flowchart

Fig.8 Target positioning

Fig.9 Angular diagram

Fig.10 Angular correction diagram

Fig.11 Y-direction pixel statistics

Fig.12 Result of segmentation

Fig.13 SE module structure

Fig.14 ECA module structure

Fig.15 Structure of the YOLOv5s-MEF model

Fig.16 Model recognition

Fig17 Work interface

Fig.18 Enhanced images

Tab.1 Experimental models under different improvement strategies

	组2	组3	组4	组5	组6
轻量化	√	√	√		√
ECA		√			√
损失函数改进			√	√	√

Tab.2 Ablation experiment results

实验组序	准确率/%	召回率/%	FPS	F1/%
1	97.7	89.2	54.64	93.25
2	96.8	88.3	65.79	92.35
3	98.2	90.1	66.23	93.97
4	97.8	90.0	61.12	93.73
5	98.3	90.3	51.62	93.98
6	99.1	92.2	62.11	95.52

Fig.19 Graph of partial processing results

Fig.20 Defect identification result

Tab.3 The result of algorithm comparisons

方法	检测时间/s	准确率/%
改进YOLOv5s模型	0.021	73
本文算法	0.083	99

Fig.21 Comparison of recognition results of different methods

Tab.4 Performance comparison of three detection algorithms

方法	检测时间/s	准确率/%
改进模板匹配法	0.286	97.7
改进SSD算法	0.052	95.5
本文算法	0.083	98.6

References 20

[1]	张德海，祝志逢，李艳芹，等.基于机器视觉的二维图像质量缺陷检测研究进展［J］.包装工程，2023，44（23）：198-207.
	ZHANG Dehai， ZHU Zhifeng， LI Yanqin，et al.Research Progress of Two-dimensional Image Quality Defect Detection Based on Machine Vision［J］.Packaging Engineering，2023，44（23）：198-207.
[2]	唐建华，张鑫，刘金海，等. 基于复合骨干网络的漏磁小缺陷信号检测方法［J］.电子测量与仪器学报，2024，38 （10）： 69-77.
	TANG Jianhua， ZHANG Xin， LIU Jinhai，et al.Signal Dtection Method for Magnetic Flux Leakage Small Defects Based on Composite Backbone Network［J］.Journal of Electronic Measurement and Instrumentation，2024，38 （10）： 69-77.
[3]	胡宜笑.印刷品缺陷检测研究与应用［D］. 长沙：湖南大学，2022.
	HU Yixiao. Research and Application of Printing Detection［D］. Changsha：Hunan University，2022.
[4]	梅文宝.基于形状模板匹配的手机商标检测技术研究［D］.广州：广东工业大学，2017.
	MEI Wenbao.The Research of Mobile Phone Logo Detection Technology Based on Shape Model Matching［D］.Guangzhou：Guangdong University of Technology， 2017.
[5]	BROWN L.A Survey of Image Registration Techniques［J］.ACM Computing Surveys，1992，24（4）：325-376
[6]	ZITOVA B， FLUSSER J. Image Registration Methods： Asurvey［J］. Image and Vision Computing，2003，21（11）： 977-1000.
[7]	LOVE D. Distinctive Image features from Scale Invariant Key-points［J］. International Journal of Vision Computer，2004，60（2）：91-110.
[8]	吴先锋.基于机器视觉的智能电表铭牌印刷缺陷检测［D］.长沙：湖南大学，2022.
	WU Xianfeng.Detection of Defects in Nameplate Printing of Smart Meter Based on Machine Vision［D］. Changsha：Hunan University，2022.
[9]	李大华，徐傲，王笋，等.基于改进YOLOv5的印刷电路板缺陷检测［J］.电子测量技术，2023，46（23）：112-119.
	LI Dahua， XU Ao， WANG Sun，et al.Printed Circuit Board Defect Detection Based on Improved YOLOv5［J］. Electronic Measurement Technology，2023，46（23）： 112-119.
[10]	童小钟，魏俊宇，苏绍璟，等.融合注意力和多尺度特征的典型水面小目标检测［J］.仪器仪表学报，2023，44（1）：212-222.
	TONG Xiaozhong， WEI Junyu， Su Shaojing，et al.Typical Small Target Detection on Water Surfaces Fusing Attention and Multi-scale Features［J］. Chinese Journal of Scientific Instrument，2023，44（1）：212-222.
[11]	楼豪杰. 基于少样本学习的印刷品微小缺陷检测方法研究［D］.西安：西安理工大学，2022.
	LOU Haojie.Reserch on Small Defect Defection Methon of Printing Matter Based on Few-shot Learning［D］. Xi'an：Xi'an University of Technology，2022.
[12]	OTSU N. A Threshold Selection Method from Gray-level Histograms［J］.IEEE Transactions on Systems Man &Cybernetics，2007，9（1）：62-66.
[13]	冯毅雄，李康杰，高一聪，等．基于特征与形貌重构的轴件表面缺陷检测方法［J］.浙江大学学报（工学版），2020，54（3）： 427-434.
	FENG Yixiong， LI Kangjie， GAO Yicong，et al. Shaft Surface Defect Detection Method Based on Feature and Morphology Reconstruction［J］. Journal of Zhejiang University（Engineering Science）， 2020， 54（3）： 427-434.
[14]	胡蝶，侯俊，张全年，等.基于卷积神经网络的生产日期识别［J］.电子测量技术，2020， 43（1）：152-156.
	HU Die， HOU Jun， ZHANG Quannian，et al. Production Date Identification Based on Convolutional Neural Network Electronic Measurement Technology，2020，43（1）：152-156.
[15]	HOWARD A， SANDLER M， CHEN B， et al. Searching for MobileNetV3 ［C］∥IEEE International Conference on Computer Vision （ICCV）.Seoul， 2020：00140.
[16]	HU J， SHEN L， SUN G，et al.Squeeze-and-Excitation Networks［C］∥IEEE Conference on Computer Vision and Pat-tern Recognition. Salt Lake City， 2018：00745.
[17]	WANG Q L， WU B G， ZHU P F， et al.ECANet：
	Efficient Channel Attention for Deep Convolutional
	Neural Networks［C］∥IEEE Conference on Computer Vision and Pattern Recognition （CVPR）.Piscataway，2020： 11531-11539.
[18]	李博宇，杨文翰，杨乐，等.基于模板匹配的肥料外包装印刷缺陷检测研究［J］.印刷与数字媒体技术研究，2023（2）：39-49.
	LI Boyu， YANG Wenhan， YANG Le，et al.Research on Defect Detection of Fertilizer Packaging Based on Template Matching［J］.Printing and Digital Media StudyTechnology，2023（2）：39-49.
[19]	尹玉梅. 基于感知哈希的图像相似检索算法研究［D］.昆明：昆明理工大学，2021.
	YIN Yumei.Research on Image Similarity Retrieval
	Algorithm Based on Perceptual Hash［D］. Kunming：
	Kunming University of Technology， 2021.
[20]	顾桂梅，王小亮.深度学习的接触网小目标缺陷识别研究［J］.电子测量与仪器学报，2024，38（4）：151-160.
	GU Jiamei， WANG Xiaoliang.Identification of the Catenary Small Target Defects in Deep Learning［J］.Journal of Electronic Measurement and Instrumentation， 2024， 38（4）：151-160.