基于工业工件识别的粗定位研究  

Preliminary Location Based on Industrial Workpiece Identification

作　　者：高宝林 万国扬 江明[1,2] 葛凌枫 GAO Baolin;WAN Guoyang;JIANG Ming;GE Lingfeng(Key Laboratory of Advanced Perception and Intelligent Control of High-end Equipment,Ministry of Education,Wuhu 241000,China;School of Electrical Engineering,Anhui Polytechnic University,Wuhu 241000,China)

机构地区：[1]高端装备先进感知与智能控制教育部重点实验室,安徽芜湖241000 [2]安徽工程大学电气工程学院,安徽芜湖241000

出　　处：《安徽工程大学学报》2023年第4期39-46,共8页Journal of Anhui Polytechnic University

摘　　要：为提升工业工件检测的识别精度和检测速率,提出一种粗定位阶段基于YOLO V3(You Only Look Once V3)算法的改进。由于工业上工件尺寸、形状差异过大等问题,YOLO V3算法在实际检测中不能很好得完成检测任务。结合评价指标和实际检测效果,通过对Anchor框进行重聚类,并构建以Efficient-Net为主干的特征提取网络,通过引入SPP池化模块,应用CIOU位置损失函数,加深网络特征提取能力的同时,扩大卷积核的感受野,实现更好的神经网络特征提取。通过实验对比验证,本文提出的改进算法的识别效果相较于YOLO V3算法有所提升,在测试集mAP达到73.3%。In order to improve the recognition accuracy and detection rate of industrial workpiece,this paper proposes an improved approach based on YOLO V3(You Only Look Once V3)algorithm in the initial positioning stage.In view of the industrial workpiece size,shape difference is too large,YOLO V3 algorithm in the actual detection can not complete the detection task.Combined with the evaluation index and actual detection effect,the Anchor box is regrouped,and the feature extraction network with efficient-net as the main backbone is constructed.By introducing SPP pooling module and applying CIOU position loss function,the feature extraction ability of the network is enhanced,and the receptive field of the convolution kernel is expanded,so as to achieve better neural network feature extraction.Through experimental comparison and verification,the recognition effect of the improved algorithm is increased compared with that of YOLO V3 algorithm,reaching 73.3%in the test set mAP.

关 键 词：深度学习 工件识别 YOLO V3 Efficient-Net CIOU 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]