- Title
- Detecting Object and Direction for Polar Electronic Components via Deep Learning
- Creator
- Chen, Wen-Shuai; Ren, Zhi-Gang; Wu, Zong-Ze; Fu, Min-Yue
- Relation
- Zidonghua Xuebao Vol. 47, Issue 7, p. 1701-1709
- Publisher Link
- http://dx.doi.org/10.16383/j.aas.c190037
- Publisher
- Zhongguo Kexueyuan Zidonghua Yanjiusuo
- Resource Type
- journal article
- Date
- 2021
- Description
- The category, direction identification and positioning of polar electronic components play an important role in the industrial production, welding and inspection. In this paper, we first successfully transform the original problem of directional identification of polar electronic components into a classification problem. Then, the Faster RCNN (region convolutional neural network) and YOLOv3 methods are used to realize the correct classification, direction identification and accurate positioning of the polar electronic components. The experiments validate the effectiveness of our proposed method and the mAP (mean average precision) of the two proposed algorithms can reach 97.05 %, 99.22 %. In addition, we improve the anchor boxes of the Faster RCNN and YOLOv3 by K-means algorithm through the length and width distributions of the target frames of the datasets, the accuracy can be improved by 1.16 %, 0.1 %. We also propose the YOLOv3-BigObject network structure for the large target detection, while improving the accuracy, the cost time for detecting a single picture is also greatly reduced. Finally, the board with the electronic components is chosen to test and good experimental results are obtained.
- Subject
- electronic manufacturing; deep learning; direction recognition; object detection; faster RCNN; citing literature
- Identifier
- http://hdl.handle.net/1959.13/1467776
- Identifier
- uon:47902
- Identifier
- ISSN:0254-4156
- Language
- Chinese
- Reviewed
- Hits: 1109
- Visitors: 1098
- Downloads: 0