智能装配中基于YOLO v3的工业零件识别算法研究  

Research on industrial parts recognition algorithm based on YOLO v3 in intelligent assembly

作　　者：张静[1] 刘凤连[1] 汪日伟[1] ZHANG Jing;LIU Feng-lian;WANG Ri-wei(Key Laboratory on Computer Vision and Systems,Ministry of Education of China,the Key Laboratory on Intelligence Computing and Novel Software Technology of the City of Tianjin,Tianjin University of Technology,Tianjin,300384,China)

机构地区：[1]天津理工大学计算机视觉与系统教育部重点实验室和天津市智能计算及软件新技术重点实验室,天津300384

出　　处：《光电子．激光》2020年第10期1054-1061,共8页Journal of Optoelectronics·Laser

基　　金：天津市教委科研重点项目(2017ZD13)资助项目。

摘　　要：传统装配系统中依靠人力进行重复性劳动,容易由于人的操作具有疲劳性和人眼分辨率有限等特点造成失误,为了避免浪费人工和时间,解决工厂环境中光线等不稳定因素,提出了一种基于YOLO v3算法对形状多样的工业零件识别方法。在智能装配系统中根据视觉检测结果判断零件种类,弥补了传统方法的不足,满足产品生产系统的节拍要求。改进后的YOLO v3网络模型使用k-means算法重新聚类预选框的参数,残差网络来减少网络的参数,结合多尺度方法、采用Mish激活函数提高精确度,使其更适合工业零件的小目标分类检测。该模型以3D打印的工业零件制作数据集,实验表明与原有的YOLO v3算法对比,使用改进后的网络模型具有良好的鲁棒性,准确率提高了1.52%,时间提高了7.25 ms,实现精确实时地检测出智能装配系统中的零件种类。The traditional assembly system relies on manual labor for repetitive labor,which is easy to cause errors due to the fatigue and limited human eye resolution of human operations.In order to avoid wasting labor,time and solve unstable factors such as light in the factory environment,it is proposed a method for recognizing industrial parts with various shapes based on YOLO v3 algorithm.In the intelligent assembly system,the types of parts are judged based on the visual inspection results,which makes up for the shortcomings of the traditional methods and meets the beat requirements of the product production system.The improved YOLO v3 network model uses the k-means algorithm to cluster the parameters of the anchor box,and the residual network to reduce the network parameters.Combined with the multi-scale method and the Mish activation function to improve accuracy,it is more suitable for small industrial parts and target classification detection.The model uses 3 D printed industrial parts to make a data set.Experiments show that compared with the original YOLO v3 algorithm,using the improved network model has good robustness,the accuracy rate is increased by 1.52%,and the time is increased by 7.25 ms.Accurately detect the types of parts in the intelligent assembly system in real time.

关 键 词：智能装配 YOLO v3 Mish 工业零件 多尺度方法 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]