Nd-Net: Exploring Non-Defective Features for More Robust Inspection of Power Transmission Line Defect

Defect detection in aerial images is a significant w ork in power transmission line inspection. Thanks to the applic ation of object detection neural network, favorable performance has been achieved in detection of common defects. However, d ue to an feature alignment problem of defective and non-defecti ve data has not been resolved in state-of-the-art research, a big gap, between performance of current models and expected accu racy in applications, still exists. In this paper, a novel network a rchitecture, consisting of a multi-label classification head follow ed by a backbone network, is proposed. The backbone network in the architecture is taking advantage of non-defective fittings features in background. Experimental results show that the pr oposed method achieves more competitively capabilities with high accuracy, compared with previous architectures.