Abstract: Since power system inspection involves many scenes (mountains,forests,rural areas,towns and so on),and the lack of power line data sets for Unmanned Aerial Vehicle (UAV) inspection,it brings challenges to the power line extraction task based on supervised learning methods.For this purpose,a set of multi-scene aerial imagery power line dataset was constructed.In view of the need to improve the accuracy of power lines extracted by existing algorithms and the requirements of mobile terminal inspection on efficiency and model size,a novel power line extraction algorithm based on the improvements of LinkNet was proposed.A lightweight network was used in the Encoder part to improve the efficiency of feature processing and reduce the size of the model.Then,an Attention Dilation block (AD-block) based on depth-wise separable convolution and channel attention mechanism was proposed to improve the receptive field of the network and enhance the feature extraction ability of the network.The Decoder part was improved by introducing the bilinear interpolation up sampling method to improve the smoothness of the detection results.The experimental results showed that the accuracy of the proposed algorithm was better than that of the comparison algorithms on the server,which had better robustness.On Jetson TX2,the accuracy and efficiency of the proposed algorithm were better than that of the comparison algorithms,and the processing speed could be achieved 115ms per image.

Key words: machine vision, deep learning, semantic segmentation, power line extraction, unmanned aerial vehicle

摘要： 由于电力系统巡检涉及山区、森林、农村和城镇等众多场景,且缺乏面向无人机巡检的电力线数据集,给基于有监督学习的电力线提取任务带来了挑战。为此,构建了一套多场景航拍电力线数据,并针对现有算法提取的电力线精度有待提升以及移动端巡检对算法效率的要求等,提出了一种新的电力线提取算法,所提算法对LinkNet框架进行改进,首先,在其Encoder部分使用轻量型网络,提高算法特征处理的效率,并降低模型的大小。然后,提出了基于深度可分离卷积和通道注意力机制的AD-block来提高网络的感受野,增强网络的特征提取能力。最后,通过引入双线性插值上采样方法等改进Decoder部分。实验结果表明,所提算法在服务器上精度优于对比算法,具有较好的鲁棒性。在Jetson TX2上验证所提算法的精度和效率优于对比算法,能实现每张图像115 ms的处理速度。

关键词: 机器视觉, 深度学习, 语义分割, 电力线提取, 无人机