基于深度学习的先进陶瓷零件实时缺陷检测系统  

Real-Time Defect Detection System for Advanced Ceramic Parts Based on Deep Learning

作　　者：马晨凯 吴毅慧 傅华奇 业宁[1] MA Chenkai;WU Yihui;FU Huaqi;YE Ning(College of Information Science and Technology,Nanjing Foresty University,Nanjing 210037,China;College of Automotive and Transportation Engineering,Nanjing Foresty University,Nanjing 210037,China;College of Electrical and Information Engineering,Heilongjiang Institute of Technology,Harbin 150050,China)

机构地区：[1]南京林业大学信息科学技术学院,南京210037 [2]南京林业大学汽车与交通工程学院,南京210037 [3]黑龙江工程学院电气与信息工程学院,哈尔滨150050

出　　处：《南京航空航天大学学报》2021年第5期726-734,共9页Journal of Nanjing University of Aeronautics & Astronautics

摘　　要：传统先进陶瓷零件检测与分类的主流方法为纯机械尺寸过滤和人工判断,为解决其成本高、失误率高和损坏率高等问题,提出了基于深度学习的多目标实时检测分类模型(Multi-object real-time detection and classification model,MRDC)。该模型以YOLOv3为基础,使用SKNet作为注意力机制进行特征重构提高精确度,配合灰度图快速转化算法与跳帧检测方法提高检测速度,可实现实时缺陷检测。对实际生产中的先进陶瓷零件进行采集训练,多批次采集图像数据,每批数据含多个陶瓷零件的1000张图像,平均精确率均值达到99.19%,用先进陶瓷零件生产线视频检验,识别分类的正确率达到100%,可以保证每分钟检测450~550个零件。多目标实时检测分类模型拥有识别速度更快、识别准确率更高和零件不易损坏等优点,可极大地节约生产原料与人力成本,减少废品产出。The mainstream detection and classification methods of traditional advanced ceramic parts are pure mechanical size filtering and manual judgment.To solve the problems of high cost,high error rate and high damage rate,a multi-object real-time detection and classification model(MRDC)based on deep learning is proposed.The model is based on YOLOv3,uses SKNet as an attention mechanism for feature reconstruction to improve accuracy,and cooperates with grayscale map fast transformation algorithm and frame skipping detection method to improve detection speed,which can realize real-time defect detection.Image data in multiple batches are collected,and each batch of data contains 1000 images of multiple ceramic parts.The mean average precision reaches 99.19%when the advanced ceramic parts in actual production are collected and trained.The correct rate of recognition and classification reaches 100%when the production line video of advanced ceramic parts is used for inspection,which can guarantee to detect 450—550 objects per minute.The MRDC model has the advantages of faster recognition,higher recognition accuracy,and less damage to objects,which can greatly save production raw materials and labor costs,reduce scrap output,and protect the environment more.

关 键 词：先进陶瓷材料 目标检测 机器学习 视频图像识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]