摘要
针对食品饮料等复杂包装上喷码质量检测的准确率不高与速度慢等问题,提出了一种基于Ghost-YOLO轻量化网络与嵌入式平台的喷码质量检测方法。网络以YOLOv5为基础,采用了幻影模块(GM)对卷积层进行降维,模型参数减少25%。多分类目标检测任务的后处理采用位置重复抑制(PDS)方法,通过对所有类别同时采用非极大值抑制(NMS),进一步提高检测精度。最后,利用所提出的改进自训练方法对模型进行训练,并将所提检测方法部署于嵌入式设备中,实现了对喷码质量的实时检测。实验结果表明,所提检测方法在满足实时性的要求下,对喷码字符检测的精确度和召回率分别达到了100%和99.99%。
Aiming at the shortcomings such as low accuracy and slow speed of quality detection for inkjet codes on the surface of complex packaging in food,beverage and other industries,a detection method for inkjet code quality based on Ghost-YOLO lightweight network and embedded platform is proposed in this paper.The network is originated from YOLOv5.Firstly,Ghost module(GM)is used to shrink the dimension of the convolutional layer,and the model parameters are reduced by 25%.Secondly,the position duplication suppression(PDS)method is used in the post-processing of multi-category object detection task.The detection accuracy is further improved by applying the non-maximum suppression(NMS)to all categories simultaneously.Finally,the proposed improved self-training method is used to train the model,and the proposed detection method is deployed in the embedded device to realize the real-time detection of inkjet code quality.The experimental results show that the precision and recall for code characters reach 100%and 99.99%,respectively.
作者
葛俏
梁桥康
邹坤霖
孙炜
李珊红
王耀南
GE Qiao;LIANG Qiao-kang;ZOU Kun-lin;SUN Wei;LI Shan-hong;WANG Yao-nan(College of Electrical and Information Engineering,Hunan University,Changsha 410082,China;National Engineering Research Center for Robot Visual Perception and Control Technology,Hunan University,Changsha 410082,China;School of Advanced Manufacturing Engineering,Hefei University,Hefei 230601,China)
出处
《控制工程》
CSCD
北大核心
2022年第12期2349-2356,共8页
Control Engineering of China
基金
国家自然科学基金资助项目(62073129,U21A20490)
湖南省自然科学基金资助项目(2022JJ10020)
安徽省自然科学基金资助项目(1808085QF195)。
