摘要
针对铸件X射线图像获取困难、人工及传统识片方法效率低且漏判率高的问题,文中提出了一种基于深度学习的铸件缺陷检测方法。首先,采用Overlap切图(重叠切图)数据增广方法实现缺陷扩充,并基于简化Mosaic数据增广进一步提升图像的复杂度;然后,基于仅浏览一次(YouOnlyLookOnce,YOLO)的理念实现缺陷检测模型构建;最后,提出一种基于边界框抑制的测试图像缺陷检测方法,以子图迭代方式完成测试图像中的缺陷检测。实验结果表明,该方法能够有效实现多种铸件缺陷的自动检测,为铸件缺陷检测提供了基于深度学习的解决方案。
Aiming at the problems of difficulty in obtaining X-ray images of castings,low efficiency of manual and traditional image recognition methods and high rate of missed judgments,a casting defect detection method is proposed in this paper based on deep learning.Firstly,the overlap data augmentation method is used to achieve defect augmentation and the image complexity is further improved based on simplified Mosaic data augmentation;then the defect detection model is constructed based on the idea of YOLO(you only look once);finally,a test image defect detection method based on bounding box suppression is proposed,which iteratively completes the defect detection in the test image.The experimental results show that the automatic detection of various casting defects can be effectively realized with this method,which provides a solution based on deep learning to the defect detection of castings.
作者
于宏全
袁明坤
常建涛
罗坤宇
YU Hongquan;YUAN Mingkun;CHANG Jiantao;LUO Kunyu(ZTE Corporation,Xi’an 710114,China;Xidian University,Xi’an 710071,China)
出处
《电子机械工程》
2021年第6期59-64,共6页
Electro-Mechanical Engineering
基金
陕西省重点研发计划项目(2020ZDLGY07-08)
陕西省重点研发计划项目(2020ZDLGY07-04)。
