Weakly-Supervised Learning Based Augmentation of Aerial Images for Insulator Fault Recognition

Cascade detectors are widely used to detect insulators’ self-blast faults from aerial images obtained by unmanned inspection systems, in which the RPN and classifier are adopted for insulator boundary extraction and fault identification. However, classifier performance can be significantly limited by the complex background and the insufficient samples. This paper proposes an augmentation method to promote the classifier based on prior knowledge and the existing RPN dataset. Firstly, a weakly-supervised framework is established to train an RRPN from the RPN dataset. Given the Oriented Bounding Box (OBB) predicted by RRPN, the extracted insulator can be combined with the background randomly to generate new samples. The proposed method is extensively assessed through experiments and the numerical results show that the RRPN can not only provide boundaries for augmentation but also replace RPN with the same detection capability. And the classifier performance can be steadily improved without any additional annotation, e.g., the F1 score of MobileNet-Small can be improved from 91.64% to 97.25%.