A Matching-Detection-Classification Cascaded Model for Small-Size Defect Identification on Transmission Lines

The online monitoring system for transmission lines employs artificial intelligence algorithms to detect potential hazards or defects. However, factors such as the wide field of view of the camera, intricate image information, and small-sized objects lead to a relatively low recognition accuracy. Furthermore, reducing false positives and enhancing the effective alarm rate of the device represent critical challenges in system implementation. In our work, we propose a matching-detection-classification cascaded model. Initially, a matching network is utilized to ascertain whether discrepancies exist between the current monitoring image and its predecessor. Upon identifying any differences, a detection network is employed to assess equipment and defects present on the transmission line itself. Finally, for small-sized targets, we implement a multi-label classification network for precise category identification. Experimental results indicate that our algorithm demonstrates state-of-the-art (SOTA) performance on the collected dataset, and the effective alarm rate in practical application reaches <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 0. 6 \%}$</tex>.