Transmission Line Defective Bolts Detection Based on Fine-grained Recognition and Mutual Learning

Jimin Yu, Yilin Ma

Abstract: Unmanned Aerial Vehicle (UAV) inspection has replaced most manual tasks to perform transmission line image capture. However, analyzing these images still relies heavily on manual labor, which is time-consuming and prone to errors. Defect bolts detection is an important part of transmission line inspection, but due to the small size of bolts and the difficulty in defects identifying, automated transmission line inspection faces challenges. Therefore, we propose a deep neural network called FG R-CNN (Fine-Grained R-CNN) to detect defective bolts in transmission lines. First, to distinguish the bolts that are highly similar between classifications, we introduce the convolutional binary tree branch to carry out fine-grained recognition of RoI features. Secondly, we introduce classification consistency loss to constrain the differences between multiple classifiers of the tree structure to promote mutual feature learning. Finally, we add the feature pyramid network (FPN) to enable our network more suitable for small bolts detection. Our experimental results on a transmission line inspection image dataset show that our model has a 6.98 % higher bolts AP and a 12.02 % higher defect bolts AP than the commonly used object detection network Faster R-CNN.

Keywords: detective bolts detection, fine-grained recognition, mutual learning, Faster R-CNN