Hybrid Knowledge R-CNN for Transmission Line Multifitting Detection
Yongjie Zhai, Xu Yang, Qianming Wang, Zhenbing Zhao, Wenqing Zhao
Abstract
Aiming at the problems of complex background, tiny-size objects, long-tailed distribution, and so on, a hybrid knowledge region-based convolutional neural network (HK R-CNN) is proposed to detect multiple fittings in aerial images of transmission lines. First, the structure combination rules of transmission line fittings are studied, and the relationships of co-occurrence connection and spatial location between fittings are effectively extracted through a data-driven way. Second, the position-sensitive score map (PSSM) is utilized to express the immobilized connection structure of the fittings and extract their visual features. Finally, distinct relationship forms are instantiated by the integrated knowledge modules based on graph learning. The proposed model can enhance the corresponding visual features and realize fitting classification and position regression. Experimental results show that the proposed model can accurately detect the multiple fittings on transmission lines, and the detection performance for sample deficiency and tiny-size fittings is improved significantly compared to the commonly used high-performance model, Faster R-CNN.