期刊
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS
卷 17, 期 5, 页码 3450-3459出版社
IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
DOI: 10.1109/TII.2020.3013277
关键词
Defect detection; dilated convolution; industrial big data (IBD); industrial inspection application; real time
类别
资金
- Key R&D Program in Key Areas of Guangdong Province [2019B010137001, 2020B010166001]
- Industrial Internet Innovation and Development Project in 2018 [MIZ1824020]
- Guangzhou City Industrial Technology Major Research Project [201802010035]
- National Natural Science Foundation of Chongqing [cstc2019jcyj-msxmX0208]
- open research fund of National Mobile Communications Research Laboratory, Southeast University [2020D05]
This article introduces a novel defect detection method based on deep learning for digital signal processing of industrial inspection applications. By utilizing a feature collection and compression network and a Gaussian weighted pooling method, the proposed method improves both accuracy and efficiency in real-time systems, achieving mAP/AP(50) of 41.8/80.2 at 33 fps on NEU-DET.
The signal processing of industrial big data (IBD) is a challenging task, owing to the complex working scenarios and the lack of annotations. Defect detection, which is an important subject of IBD research works, has shown its effectiveness in digital signal processing of industrial inspection applications in many previous studies. This article proposes a novel defect detection method based on deep learning for digital signal processing of industrial inspection applications. In our method, a module named feature collection and compression network is applied to merge multiscale feature information. Then, a new pooling method named Gaussian weighted pooling, which provides more precise location information, is used to replace region of interest (ROI) pooling. Experiment results show that our method gets improvements in both accuracy and efficiency, with mAP/AP(50) of 41.8/80.2 at 33 fps on NEU-DET, which satisfies the requirement of real-time systems.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据