DIDNet: An End-to-End Directional Insulator Detection Network Based on Direction Field

Yuxiang Wu, Shuifa Sun, Ning Wei

Abstract: Insulator strings exhibit strong directional characteristics in UAV images and have relatively long spatial spans. Therefore, traditional anchor box-based methods struggle to accurately locate their spatial features. In this paper, we model the linear features of insulator strings by first describing the position of the central axis segment of the insulator string using endpoints at both ends. Then, a direction field on the insulator space is established to encode the repetitive texture features of the insulator string itself. Based on this model, an end-to-end convolutional neural network is constructed to generate a set of proposed line segments by estimated keypoints. These proposed lines are then encoded using a feature map enhanced with the direction field, enabling the learning of the texture features of the insulator string. Experimental results demonstrate that the proposed algorithm framework achieves more accurate directional positioning of insulator strings under the DIOU (Direction Intersection over Union) metric.