基于改进ResNet101网络的齿轮缺陷检测
Gear Defect Detection Based on Improved ResNet101 Network
包从望 1江伟 2刘永志 2肖钦兰 2吴娇2
作者信息
- 1. 六盘水师范学院矿业与机械工程学院,六盘水 553000;中国矿业大学机电工程学院,徐州 221116
- 2. 六盘水师范学院矿业与机械工程学院,六盘水 553000
- 折叠
摘要
针对视觉技术下的齿轮缺陷检测中,存在检测精度低、特征提取能力弱及检测模型不稳定等问题,提出了一种改进ResNet101 网络的齿轮缺陷检测方法.首先,基于ResNet101 网络,引入空洞卷积操作,在各个残差层中引入不同比例的膨胀系数,实现齿轮图像不同感受野下的特征提取;其次,在各个卷积模块间引入稠密连接操作,保留浅层特征信息,降低了模型训练过程中梯度消失的风险;最后,通过图像样本旋转操作,获得齿轮缺陷样本,通过准确率、召回率、ROC曲线、AUC等参数对所提方法的性能进行验证.实验结果表明,改进后的ResNet101 能有效实现齿轮缺陷检测,同时具有更高的稳定性能,可用于齿轮生产过程中,产品质量的实时在线检测.
Abstract
Aiming at the problems such as low detection accuracy,weak feature extraction ability and unsta-ble detection model,a gear defect detection method with improved ResNet101 network was proposed in this paper.Firstly,based on ResNet101 network,the cavity convolution operation was introduced,and the expan-sion coefficients of different proportions were introduced into each residual layer to realize feature extraction under different sensitivity fields of the gear image.Secondly,dense joint operation was introduced between each convolutional module to retain shallow feature information,which reduces the risk of gradient disappea-ring during model training.Finally,the gear defect samples were obtained by rotating operation,and the per-formance of the proposed method was verified by accuracy,recall rate,ROC curve,AUC and other parame-ters.The experimental results show that the improved ResNet101 can effectively detect gear defects and has higher stability performance.It can be used for real-time on-line detection of product quality in gear production.
关键词
深度学习/ResNet101/网络/齿轮缺陷/特征提取Key words
deep learning/ResNet101 network/gear defects/feature extraction引用本文复制引用
基金项目
贵州省教育厅基金项目(黔教合KY字[2020]117)
六盘水市科技计划项目(52020-2019-05-12)
六盘水市科技计划项目(52020-2022-PT-02)
六盘水师范学院基金项目(LPSSYylzy2205)
六盘水师范学院基金项目(LPSSYzyzhggsd201802)
出版年
2024
NETL
NSTL
万方数据