Exploring the Generalizability of Recent Object Detection Models in Identifying Defective Glass Insulators for UAV Power Line Inspection A Case Study in Morocco

Abstract

Power line inspection through drone technology improved by computer vision has attracted considerable investment from the electric transportation sector, by dint of its potential to enhance operational efficiency, reduce costs and attenuate risk. However, creating proprietary solutions for companies can be prohibitively costly due to the extensive research and development processes involved. That’s why many organizations opt to procure foreign solutions to expedite their integration into operational protocols. Yet, a key concern is, the adaptability and effectiveness of these foreign solutions, in new operational context: Can externally procured solutions meet the purchaser operational requirements? Our research focuses on evaluating the generalization ability of recent high-performance object detection models, specifically in identifying damaged glass power line insulators in drone-acquired images. To investigate, we carefully constructed two datasets—Vietnamize and Moroccan. We then trained and validated six recent models on the Vietnamese dataset and evaluated their performance on a new environment - a Moroccan dataset. Our findings indicate that these models consistently demonstrate high levels of performance in the Vietnamese region where they were trained. In contrast, when applied to the Moroccan context, their performance metrics experienced a decline. These results highlight the fact that current models still suffer from the generalization problem when they discover a new environment. Also, these findings emphasize that entities seeking intelligent external solutions should recognize the absence of a universally applicable solution. Instead, it becomes imperative for organizations to equip themselves with the requisite mechanisms to tailor the acquired solution effectively to its specific operational milieu.