DPNet: A Lightweight Directional-Aware Pointwise Network for Dropper Defect Detection in High-Speed Railway Preventive Maintenance

In overhead power system (OPS), the contact lines are suspended on the bearing cable through the droppers, and the structural height of the contact suspension is controlled by adjusting the length of the droppers. Therefore, carrying out automatic and reliable dropper defect detection technology is essential to guarantee the safety of railway power supplement and avert transportation accidents. However, due to arbitrary orientations and extreme aspect ratio, locating and identifying the dropper defects pose great difficulties. Most existing dropper detection approaches either focus only on detecting the presence of dropper instances without further distinguishing specific defect states or have limited exploration to analyze dropper defect categories. In this article, we propose a directional-aware pointwise network (DPNet), which contains a depthwise-separable multidirectional convolution (DMC) to perceive different rotated dropper instances. Compared with traditional convolution operations, our DMC is more lightweight while solves the problem of precision degradation caused by the original depthwise-separable convolution (DSC), which is more suitable to be deployed in our detection machine. Moreover, to alleviate the problem of instance ambiguity caused by the extremely similar appearance, a new pointwise localization detector (PLD), which predicts start, center, and endpoint of droppers, is introduced to refine the final detection results. We compare our proposed DPNet with many state-of-the-art approaches. Extensive experimental comparisons demonstrate that our method successfully detects most dropper defects, which achieves more precise performance compared with all other detection methods by a large margin.