关键词: 关键词: YOLOv11n, 输电线路巡检, 特征提取, 部分卷积, 注意力机制, 轻量化

Abstract: Abstract: To address the issues of low efficiency, high misdetection rates and high omission rates in traditional power line inspections, an improved model based on YOLOv11n, named ETL-YOLO, was proposed. Firstly, to address the limited computational efficiency and insufficient multi-scale feature representation of conventional bottleneck designs, a lightweight multi-scale bottleneck structure, MSLW Bottleneck, was introduced. This structure was integrated into the C3k2 module to form the C3k2-MSLW feature extraction unit, enabling enhanced multi-scale representation while reducing both parameter count and computational overhead. Secondly, the loss of fine details caused by conventional downsampling convolutions was mitigated through the incorporation of the PEConv module, which leveraged the lightweight characteristics of PConv together with the multi-scale feature learning capability of EMA to enhance the detection of small objects. Finally, difficulties in identifying small targets under complex backgrounds were alleviated through the introduction of the PECAA attention mechanism, enabling selective enhancement or suppression of feature responses and thereby improving sensitivity to small objects in challenging environments. Experimental results on the public InsPLAD dataset show that the improved model achieves accuracy, recall, and mAP50 values of 93.5%, 92.6%, and 96.2%, with gains of 2.9%, 4.2%, and 1.7% over the baseline, respectively. The parameter count, computational complexity, and weight file size are reduced by 43.6%, 10.9%, and 40.3%. Overall, the model delivers a synergistic improvement in detection accuracy and efficiency and serves as an effective solution for UAV-based power line inspection.

Key words: Keywords: YOLOv11n, transmission line inspection, feature extraction, partial convolution, attention mechanisms, lightweight