Few-Shot Object Detection in Unmanned Aerial Vehicles-Based Transmission Line Inspection: A Method Based on Transfer Learning and Attention Mechanism

During the operation of utilizing unmanned aerial vehicles (UAVs) to inspect transmission lines, insufficient samples and complex backgrounds may lead to poor detection quality of faults, which affects the normal operation of the power system. Therefore, a multiscale fault object detection network based on transfer learning and attention mechanism (TLAM-Det) is designed, which includes a feature preextraction network (FPENet) and a fault object detection network based on sparse feature extraction and attention mechanism (SAMNet). Specifically, a multiscale feature extraction module (INR-Block) is designed to construct FPENet, which effectively prelearns fault features from the source data. Then, to solve the low precision caused by few-shot data, the transfer learning are utilized to transfer the pretrained weights and parameters to SAMNet, which aims to improve the feature extraction capability. A spatial feature extraction and fusion module (SAMF) based on attention mechanism is designed to enhance the feature compression capability and resist complex background interference. Finally, extensive experiments were conducted to evaluate the detection performance of TLAM-Det. The experimental results were compared with ten state-of-the-art (SOTA) detection methods and achieved the highest mean average precision (mAP) of 94.4%, which is an average increase of 9.77%.