【摘要】 针对传统目标检测算法检测精度低、速度慢的问题，提出一种改进YOLOv4的目标检测模型，对架空线路中的电杆、变压器、电杆倾斜以及绝缘子跌落4类常见的电力设备及故障进行检测。模型采用针对嵌入式平台设计的MobileNet代替原YOLOv4中的骨干网络，使模型轻量化，为进一步降低模型的运算复杂度，在其颈部网络引入深度可分离卷积，同时为了加强卷积神经网络的学习能力，在颈部网络中使用了跨阶段局部网络（cross stage partial networks,CSPNet）结构。利用改进后的模型对架空线路图像数据集进行实验，实验结果表明，该模型能够在检测精度相当的情况下将检测速度提升为原模型的1.68倍，能够更好地应用到嵌入式设备上，利用无人机实现对架空线路中常见的电力设备及故障进行实时的检测。更多还原

【Abstract】 Aiming at the problems of low detection accuracy and slow speed of traditional target detection algorithms, an improved YOLOv4 target detection model is proposed to detect four types of common power equipment and faults in overhead lines, such as poles, transformers, pole tilts, and insulators drop. Instead of the backbone network in the original YOLOv4, MobileNet which is designed for the embedded platform is deployed in this model, making this model lightweight. In order to further reduce the computational complexity and strengthen the learning ability of the convolutional neural network, a deep separable convolution and a CSP structure is introduced in the neck network. This improved model is used to conduct experiments on the overhead line image data set, and the experimental results show that this model can increase the detection speed to 1.68 times of the original model with a equivalent detection accuracy. It can be better applied to embedded devices, and thus achieves the real-time detection of common power equipment and faults in overhead lines by drones.更多还原