基于注意力特征融合YOLOv5模型的无人机输电线路航拍图像金具检测方法

无人机采集输电线路航拍图像由于其特殊性，往往背景复杂多变，检测目标存在尺度不一及部分遮挡等问题容易造成检测过程中误检、漏检。本文从特征融合角度出发，提出基于注意力特征融合YOLOv5模型的输电线路金具检测方法。首先，在主干提取网络中引入了具有自注意力机制的AFF-Transformer模块更好的捕获全局信息和上下文信息，提高主干网络特征提取能力。其次，通过在特征融合过程中使用通道空间注意力避免了关键信息丢失。最后，利用双向加权特征融合机制使得模型更有效的将浅层特征和深层特征进行融合，以上改进有效缓解了金具在密集状态下的误检、漏检等问题。通过在自建输电线路金具数据集上进行实验，结果表明：本文提出的方法在原YOLOv5模型的基础上准确率提升了2.7%，模型召回率提高了1.5%，针对于小目标，以及漏检、误检等问题有了较好的改善。

Transmission line image fitting detection method based on attention feature fusion YOLOv5 model

Due to its particularity, the aerial images of transmission lines collected by UAVs often have complex and changeable backgrounds, and the detection targets have problems such as different scales and partial occlusions, which may easily lead to false detections and missed detections in the detection process. From the perspective of feature fusion, this paper proposes a transmission line fitting detection method based on the attention feature fusion YOLOv5 model. First, the AFF-Transformer module with self-attention mechanism is introduced into the backbone extraction network to better capture global information and context information, and improve the network. Feature extraction capability. Second, by using channel spatial attention, key information loss during feature fusion is avoided. Finally, the two-way weighted feature fusion mechanism is used to make the model more effectively fuse the shallow features and deep features, which effectively alleviates the problems of false detection and missed detection in the dense state. Through experiments on the data set of self-built transmission line fittings, the results show that the method proposed in this paper improves the accuracy rate by 2.7% and the model recall rate by 1.5% based on the original YOLOv5 model. Problems such as false detection have been better improved.