摘要
本文对带钢表面斑块、裂纹、夹杂等6种缺陷进行研究,利用DenseNet深度学习网络和PyQt5设计一种缺陷智能识别系统,可以实现带钢表面缺陷准确高效的识别。该识别系统在以Tensorflow为后端的Keras平台上搭建,采用迁移学习的方法对带钢表面的6种缺陷进行识别,训练过程中冻结基础模型DenseNet的顶层部分,利用数据扩充、添加BN层防止过拟合。最终模型在训练集上的正确率为99.33%,在测试集上每一类缺陷的正确率均超过97%,其间绘制出混淆矩阵。最后,搭建缺陷识别系统的GUI界面,实现带钢表面缺陷识别的可视化功能,提高用户体验。
In this paper,six kinds of defects such as patches,cracks,and inclusions on the surface of strip steel were studied,and an intelligent defect recognition system was designed by using DenseNet deep learning network and PyQt5,which could realize accurate and efficient identification of surface defects of strip steel.The recognition system was built on the Keras platform with Tensorflow as the backend,which used the transfer learning method to identify the six defects on the surface of the strip,froze the top part of the basic model DenseNet during the training process,used data to expand and added a BN layer to prevent overfitting.The final model had a correct rate of 99.33% on the training set,and the correct rate of each type of defect on the test set exceeded 97%,during which a confusion matrix was drawn.Finally,the GUI interface of the defect recognition system was built to realize the visual function of surface defect recognition of strip steel and improve the user experience.
作者
郝用兴
庞永辉
HAO Yongxing;PANG Yonghui(School of Mechanical Engineering,North China University of Water Resources and Electric Power,Zhengzhou Henan 450045)
出处
《河南科技》
2021年第3期11-14,共4页
Henan Science and Technology
