基于深度学习的子弹缺陷检测方法  

Bullet Defect Detection Method Based on Deep Learning

作　　者：利照坚 江秀娟[1] 朱铮涛 袁浩期 LI Zhao-jian;JIANG Xiu-juan;ZHU Zheng-tao;YUAN Hao-qi(School of Electromechanical Engineering,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学机电工程学院

出　　处：《组合机床与自动化加工技术》2019年第9期102-106,110,共6页Modular Machine Tool & Automatic Manufacturing Technique

摘　　要：针对子弹的表面缺陷检测使用常规的图像处理效果并不理想,文章提出使用改进的卷积神经网络和FasterRCNN网络相结合的方法,首先采集子弹不同类型的缺陷图像作为数据集,然后使用改进后的VGG16和ResNet50两种卷积神经网络分别处理数据集图像,再经过RPN网络和FastRCNN网络的训练,对图像进行分类与缺陷特征提取,获取比较准确的缺陷区域的框图。实验结果表明,使用改进后的卷积神经网络与FasterRCNN网络相结合的方法可以提高子弹缺陷检测的准确率、召回率和mAP,准确率、召回率和mAP均可达到90%。Aiming at the unsatisfactory effect of conventional image processing in bullet surface defect detection,this paper proposes a method combining improved convolution neural network with Faster RCNN network.Firstly,different kinds of defect images of bullets are collected as data sets,and then two kinds of convolution neural networks,improved VGG16 and ResNet50,are used to segment them.The data set images are processed separately,and then trained by RPN network and Fast RCNN network,image classification and defect feature extraction are carried out to obtain more accurate defect region block diagram.The experimental results show that the combination of convolutional neural network and Faster RCNN network can improve the accuracy and recall rate of bullet defect detection,and the accuracy,recall rate and mAP can reach 90%.

关 键 词：卷积神经网络 目标识别 子弹检测 准确率 召回率 

分 类 号：TH16[机械工程—机械制造及自动化] TG506[金属学及工艺—金属切削加工及机床]