Master of Science Duy Khoi Tran will Monday November 10th, 2025, at 12:15 hold his Thesis Defense for the PhD degree in Science. The title of the thesis is:
« Towards Data-efficient Fully Automated Visual Inspection of Power-Line Systems »
Keeping the electricity running to power society is important and requires thorough and regular checks of the power line infrastructure. However, conventional approaches to this are slow, costly and hazardous for people. This thesis introduces artificial intelligence solution that help automate these inspection tasks. The proposed solutions do not rely on the large amount of labeled data, which these types of solutions normally need. In particular, the thesis makes three contributions: LSNetv2, a solution to find thin power lines in images produced using sketch labels; WOODWORK, a solution to detect and measure damage caused by woodpeckers on utility poles without pixel-perfect labels; and SAM-ReID, a solution to find the same insulator across images without intricate information. Together these methods are a step forward in the quest to further automate inspection of power line systems, making it better, cheaper and safer to ensure reliable electricity. Methods of inspection traditionally demand the involvement of human personnel in both data acquisition and analysis steps. Observations/data acquisition is typically performed by inspectors approaching the systems on foot, by climbing, or using low-flying helicopters. Data analysis also requires inspectors with expertise to examine the observations on-site or in imagery captured beforehand. This conventional inspection approach is slow, expensive and potentially dangerous for the humans involved. There have been continual efforts to automate inspections, yet due to the complexity and demand for accuracy of many of the inspection tasks, achieving generalizable and fully automated power line inspection remains difficult. Motivated by that fact, this dissertation was carried out to investigate the potential of Deep Learning, with a focus on computer vision, in further advancing the viability of automating vision-based power line inspection. Specifically, the studies in this thesis tackle two identified challenges that hinder the adoption of deep learning in this domain. First is the limitation of suitably labeled data needed to train models. This is a consistent challenge of deep learning as a whole, since state-of-the-art solutions often rely on large amounts of annotated images to fully realize their capabilities. Obtaining this large volume is non-trivial as the time and resource cost of annotation can be immense. This problem is exacerbated in the power line visual inspection domain due to complex characteristics of visual inspection imagery and an increased need for more meticulous annotation because of the domain's high accuracy demands. The second challenge is the research gap that hinders the bridge between deep learning and the inspection domain. Prominent deep learning methods might not be readily suited for particular inspection tasks (slender power line detection, cross-view re-identification of small components) and certain detection targets (components or defects) can be underrepresented due to specific attributes. We contributed to alleviating these challenges by proposing three novel deep learning methods tackling three power line visual inspection tasks. All proposed methods were designed to perform well under the limitation of annotated data. LSNetv2, a novel power line detection model, was trained with weak polyline supervision, which is significantly more accessible than the granular pixel-level ground truth. WOODWORK, a woodpecker damage estimation framework, requires only image-level and bounding-box annotations for segmentation throughout its pipeline. Finally, SAM-ReID is a multi-view component re-identification solution that does not rely on metadata such as GPS and camera intrinsics and is effective when trained with a relatively small amount of data. The proposed methods were evaluated comprehensively via experiments and the concrete results obtained corroborate their effectiveness.
1st Opponent: Lecturer Aiden Durrant, University of Aberdeen, Scotland
2nd Opponent: Professor Reza Arghandeh, Western Norway University of Applied Sciences, Bergen
Internal member and leader of the committee: Associate Professor Elisabeth Wetzer, Department of Physics and Technology, UiT
The defence and trial lecture will be streamed from these following links at Panopto:
Defence (12:15 - 15:00)
Trial Lecture (10:15 - 11:15)
The thesis is available Here