摘要
基于传统X射线图像的铝合金轮毂铸件缺陷检测方法存在人工检测效率低、误检率高、检测精度较差等问题,提出一种基于深度学习的铝合金轮毂铸件图像缺陷检测方法。通过引入直方图均衡化方法,实现533组铝合金铸件X射线图像缺陷特征增强;同时基于Mosaic数据增广策略随机生成含有多尺度不同缺陷类型的新图像数据,提升图像的复杂度;修改了YOLOv5主干网络,引入SENet注意力机制模块对输入特征图的重要通道进行特征提取增强。结果表明,该方法对铸件缺陷平均检测精度(mAP)达到了99.6%,对比YOLOv3、YOLOv4以及YOLOv5主流算法,平均检测精度分别提升了9%、5.1%、4.2%。相较于原网络模型,常见的4种类型(气孔、缩松、裂纹、夹杂)铸件缺陷平均检测精度提升了10.83%。该方法具有更好的泛化能力,可实现铸件多类型缺陷的自动检测,能够满足工业实际需求。
Considering the existing problems of traditional X-ray image defect detection method for aluminum alloy hub castings,such as low manual detection efficiency,high error detection rate,and poor detection accuracy,a new method for aluminum alloy hub castings was proposed based on deep learning.By introducing the histogram equalization method,defect feature enhancement of 533 groups of X-ray images of aluminum alloy castings were achieved.Meanwhile,new image data with multi-scale and different defect types were generated randomly based on mosaic data augmentation strategy to improve the complexity of the image.The YOLOv5 backbone network was modified,and the SENet attention mechanism module was introduced to enhance the feature extraction of the important channels of the input feature map.The results indicate that the average detection accuracy(mAP)for casting defects reaches 99.6%,which is improved by 9%,5.1%and 4.2%,respectively,compared with the mainstream algorithms of YOLOv3,YOLOv4 and YOLOv5.Compared with the original network model,the average detection accuracy of the four common types of casting defects(blowhole,porosity,crack and slag inclusion)is improved by 1o.83%.The method possesses desirable generalization ability,which can realize the automatic detection of multiple types of defects in castings,and meet the actual demand of industry.
作者
闫学顺
汪东红
吴文云
姜淼
邱慧慧
龚潜海
疏达
Yan Xueshun;Wang Donghong;Wu Wenyun;Jiang Miao;Qiu Huihui;Gong Qianhai;Shu Da(School of Materials Science and Engineering,Shanghai University of Engineering Science;School of Materials Science and Engineering,Shanghai Jiaotong University;Jiashan Xinhai Pecision Casting Co.,Ltd.;Huzhou Dingsheng Machinery Technology Co.,Ltd.;Zhejiang Jiali Wind Energy Technology Co.,Ltd.)
出处
《特种铸造及有色合金》
CAS
北大核心
2023年第4期457-463,共7页
Special Casting & Nonferrous Alloys
基金
国家重点研发计划资助项目(2020YFB1710101,2022YFB3706800)
国家科技重大专项资助项目(J2019-Ⅵ-0004-0117)
国家自然科学基金资助项目(51821001,52074183,52090042)
浙江省重点研发计划资助项目(2020C01056,2021C01157,2022C01147)
长寿命高温材料国家重点实验室开放基金资助项目(DECSKL202109)。
关键词
铝合金铸件
缺陷检测
深度学习
X射线图像
注意力机制
Aluminum Alloy Castings
Defect Detection
Deep Learning
X-ray Image
Attention Mechanism
