FAFNet: An Overhead Transmission Line Component Detection Method Based on Feature Alignment and Fusion

To ensure the stability of the power supply, the paradigm of using unmanned aerial vehicles (UAVs) to capture images of overhead lines and then using visual recognition technology to identify components and their defects has become widespread. However, significant challenges remain in identifying components with different scales and high aspect ratios from complex backgrounds. From the perspective of feature representation, this paper proposes FAFNet, which includes two novel, simple, and effective modules: the feature alignment (FA) module and the light-weight shared convolutional image pyramid network (LSCIPN) module. FA solves the problem of feature misalignment when merging higher and lower level feature maps, and LSCIPN enhances the low-level detail information of the feature pyramid. The two modules work together to improve the multi-scale detection capability. By conducting extensive experiments on three datasets, namely CPLID, OTLID (constructed in this paper), and IDID, as well as on the VOC and COCO datasets for generic scenarios, we demonstrate the effectiveness and generalization performance of the proposed method. Concretely, FAFNet improved 3.3 mAP, 1.2 mAP, 5.7 mAP, 1.1 mAP, and 1.5 mAP in CPLID, OTLID, IDID, VOC, and COCO, respectively, compared to the baseline model. We hope that FAFNet could serve as an alternative solution for UAV inspection visual recognition systems in the power industry.