Lightweight Power-Line Visual Detection in Agricultural UAV Scenarios Based on an Improved YOLOv12n Model

To address the problems of low detection accuracy, slow inference speed, and high computational cost in power-line detection during autonomous operations of agricultural UAVs, this study proposes an improved object detection model based on YOLOv12n. A power-line dataset was constructed using real-field images supplemented with the TTPLA dataset. The lightweight EfficientNetV2 was introduced as the backbone network to replace the original backbone. In the neck, dynamic snake convolution and a multi-scale cross-axis attention mechanism were incorporated, while the region attention partitioning and residual efficient layer aggregation network from the baseline model were retained. In the head, a Mixture of Experts (MoE) layer from ParameterNet was integrated. The improved model achieved 80.07%, 43.07%, and 77.35% of the original model’s parameters, computation, and weight size, respectively. With an IoU threshold greater than 0.5, the mean average precision (mAP0.5) reached 75.5%, representing improvements of 13.53%, 15.09%, 7.5% and 7.54%over YOLOv8n, YOLOv11n, YOLOv5n, and Line-YOLO, respectively. Only inferior to RF-DETR-Nano. On mobile-end testing, the inference speed reached 88.36 FPS and exhibits the highest inference speed across all experimental models. The improved model demonstrates excellent generalization, robustness, detection accuracy, target localization, and processing speed, making it highly suitable for power-line detection in agricultural UAV applications and providing technical support for future autonomous and intelligent agricultural operations.