Institutional Repository
Technical University of Crete
Technical University of Crete
EN
|
EL
| URI: | http://purl.tuc.gr/dl/dias/10B27BF2-3D54-42DA-B8C3-37239BBA81D6 | ||
| Year | 2022 | ||
| Type of Item | Conference Full Paper | ||
| License |
|
||
| Bibliographic Citation | A. Tsellou, K. Moirogiorgou, G. Plokamakis, G. Livanos, K. Kalaitzakis and M. Zervakis, "Aerial video inspection of Greek power lines structures using machine learning techniques," in Proceedings of the 2022 IEEE International Conference on Imaging Systems and Techniques (IST 2022), Kaohsiung, Taiwan, 2022, doi: 10.1109/IST55454.2022.9827761. https://doi.org/10.1109/IST55454.2022.9827761 | ||
| Appears in Collections |
Power line inspection is a crucial task for the uninterrupted operation of an electricity distribution network. Till date, it is mainly carried out using manned helicopters or foot patrol. However, autonomous, intelligent inspection using unmanned aerial vehicles (UAVs) equipped with camera sensors has come to the fore lately as it can offer an advantageous automated way to deliver the task of inspection. For the accurate detection of the power lines in the imagery acquired, different state-of-the-art semantic segmentation techniques have been used. In this work, attention is mainly paid to the structure of the power lines, in order to find a proper deep learning architecture that can segment them efficiently, preserving their thin shape and reducing background noise. It is found out that DNNs that employ dilated convolutions can reach this goal and achieve high performance. The architectures in this work were evaluated in both literature datasets and videos collected by HEDNO S.A. (Hellenic Electricity Distribution Network Operator S.A.) using UAVs. Results show that, out of the four deep learning-based segmentation architectures used in the experiments, the D-LinkNet architecture, first introduced for road segmentation purposes in high-resolution satellite imagery, outperformed the others in terms of F'l-Score in various background scenarios.
Copyright © DIAS 2013
