Symbol Detection in a Multi-class Dataset Based on Single Line Diagrams using Deep Learning Models
Issued Date
2023-01-01
Resource Type
ISSN
2158107X
eISSN
21565570
Scopus ID
2-s2.0-85170639177
Journal Title
International Journal of Advanced Computer Science and Applications
Volume
14
Issue
8
Start Page
43
End Page
56
Rights Holder(s)
SCOPUS
Bibliographic Citation
International Journal of Advanced Computer Science and Applications Vol.14 No.8 (2023) , 43-56
Suggested Citation
Bhanbhro H., Hooi Y.K., Kusakunniran W., Amur Z.H. Symbol Detection in a Multi-class Dataset Based on Single Line Diagrams using Deep Learning Models. International Journal of Advanced Computer Science and Applications Vol.14 No.8 (2023) , 43-56. 56. doi:10.14569/IJACSA.2023.0140806 Retrieved from: https://repository.li.mahidol.ac.th/handle/123456789/90078
Title
Symbol Detection in a Multi-class Dataset Based on Single Line Diagrams using Deep Learning Models
Author(s)
Author's Affiliation
Other Contributor(s)
Abstract
Single Line Diagrams (SLDs) are used in electrical power distribution systems. These diagrams are crucial to engineers during the installation, maintenance, and inspection phases. For the digital interpretation of these documents, deep learning-based object detection methods can be utilized. However, there is a lack of efforts made to digitize the SLDs using deep learning methods, which is due to the class-imbalance problem of these technical drawings. In this paper, a method to address this challenge is proposed. First, we use the latest variant of You Look Only Once (YOLO), YOLO v8 to localize and detect the symbols present in the single-line diagrams. Our experiments determine that the accuracy of symbol detection based on YOLO v8 is almost 95%, which is more satisfactory than its previous versions. Secondly, we use a synthetic dataset generated using multi-fake class generative adversarial network (MFCGAN) and create fake classes to cope with the class imbalance problem. The images generated using the GAN are then combined with the original images to create an augmented dataset, and YOLO v5 is used for the classification of the augmented dataset. The experiments reveal that the GAN model had the capability to learn properly from a small number of complex diagrams. The detection results show that the accuracy of YOLO v5 is more than 96.3%, which is higher than the YOLO v8 accuracy. After analyzing the experiment results, we might deduce that creating multiple fake classes improved the classification of engineering symbols in SLDs.