针对输电线路无人机实时巡检过程中,通用目标检测算法在移动端运行速度过慢或无法运行的问题,提出一种将多尺度特征融合方法与输电线路关键部件的检测相结合的算法MSFF-KCD(Multi-scale feature fusion in key component detection)。该算法结合关键部件的特征,使用深度可分离卷积设计了特征提取网络DPNets,提高了算法在具有边缘计算能力的移动端ARM设备上的运行速度,同时采用多尺度特征融合方法,将分辨率低的特征图与分辨率高的特征图进行融合,使用多个特征融合后的特征图进行分类与检测,提高了算法的平均精度。选取了绝缘子、悬垂线夹、防震锤三类关键部件进行实验,结果表明,该算法在ARM设备上可达到每张66ms的检测速度和86%的准确率,适用于移动端关键部件检测。
英文摘要:
【】Aiming at the problem that the general target detection algorithm runs too slowly or cannot run on the mobile terminal during the real-time inspection of the UAVs on the transmission line, an algorithm MSFF-KCD combining the multi-scale feature fusion method with the detection of key components of the transmission line is proposed. The algorithm combines the characteristics of key components, and uses the depthwise separable convolutions to design the feature extraction network DPNets, which improves the running speed of the algorithm on the mobile ARM device with edge computing capability, and adopts the multi-scale feature fusion method. The low-resolution feature maps are combined with high-resolution feature maps, and the feature maps with multiple feature fusions are used for classification and detection, which improves the average accuracy of the algorithm. Three key components of insulator, suspension clamp and anti-vibration hammer are selected for experiments. The results show that the proposed algorithm can achieve a detection speed of 66ms and an accuracy of 86.09% on ARM devices, which is suitable for mobile terminal key components detection.