摘要：

为了解决复杂环境下电力巡检无人机对输电线路鸟巢识别精度低、错检漏检率高、定位不准等问题, 在YOLOv5s模型的基础上, 提出一种用于输电线路鸟巢目标检测的改进YOLO-nc-kd模型。设计一种高效的多尺度卷积特征融合模块(MCFFM), 实现不同尺度下的高效特征提取, 使模型能获得更加丰富和多样化的特征表示。引入注意力机制, 提升主干网络在相似环境背景下的鸟巢特征提取能力。设计改进的定位损失函数, 提高边界框的定位精度和小目标检测能力。使用知识蒸馏技术, 进一步提升模型精度。实验结果表明, 改进YOLO-nc-kd模型的准确率、召回率以及平均精度均值(mAP)相较于YOLOv5s模型分别提升了7.3、5.6、4.9个百分点, 具有较好的输电线路鸟巢目标图像检测效果。

关键词: 鸟巢检测, YOLOv5s模型, 注意力机制, 损失函数, 知识蒸馏

Abstract:

To solve the problems of low accuracy, high false and missed detection rates, and the inaccurate positioning of birds' nests by power inspection drones in complex environments, an improved YOLO-nc-kd model, designed to detect birds' nests on transmission lines, is proposed herein based on the YOLOv5s model. An efficient Multi-scale Convolutional Feature Fusion Module (MCFFM) is designed to achieve efficient feature extraction at different scales, enabling the model to obtain richer and more diverse feature representations. An attention mechanism is introduced to enhance the ability of the backbone network on extracting features of the bird's nest in similar environmental backgrounds. A new localization loss function is designed to improve the localization accuracy and small object detection capability of bounding boxes. Knowledge distillation techniques are implemented to further improve the model accuracy. According to the experimental results, the accuracy, recall, and mean Average Precision (mAP) of the proposed YOLO-nc-kd model are improved by 7.3, 5.6, and 4.9 percentage points, respectively, compared to those of the YOLOv5s model, indicating good detection performance for birds' nests on target images of transmission lines.

Key words: bird's nest detection, YOLOv5s model, attention mechanism, loss function, knowledge distillation