Abstract: In order to solve the obstacle avoidance issue of UAV power patrol inspection in complex environment, a UAV heading recognition method based on Inception-Resnet-V2 network is studied and improved. The depth separable convolution is introduced, and the convolutional operation is decomposed into two processes：depth convolution and point-by-point convolution, which compresses the amount of calculation. The improved network structure ensures high-precision recognition and saves the calculation cost. The improved network model achieves an accuracy of 92.5% on the standard data set. In the actual power inspection experiment, the heading prediction accuracy of the improved network model for the base tower is 95.63%. Experimental results show that UAV equipped with the improved Inception-Resnet-V2 network model can successfully identify large base towers and accurately identify and predict the heading in complex environments.

Key words: image recognition, convolutional neural network, separable convolution, course identification, Inception-Resnet-V2 network

摘要： 为解决无人机在复杂环境下电力巡检的避障难题，研究并改进了基于Inception-Resnet-V2网络的一种无人机航向识别方法。引入深度可分离卷积，将卷积操作分解为深度卷积和逐点卷积两个过程，压缩了计算量。改进后的网络结构保证高精度的识别，同时节约了计算成本。改进后的网络模型在标准数据集上达到了92.5%的准确率。在实际电力巡检实验中，改进的网络模型针对于基杆塔的航向预测精度达到95.63%。实验结果表明，搭载改进后Inception-Resnet-V2网络模型的无人机可以在复杂环境下成功识别大型基杆塔并进行精确地航向识别与预测。

关键词: 图像识别, 卷积神经网络, 可分离卷积, 航向识别, Inception-Resnet-V2网络