基于无人机巡检图像的绝缘子串实时定位研究

Research on Real-Time Positioning of Insulator Strings Based on UAV Inspection Images

针对无人机巡检输电线路图像中存在绝缘子串定位难点,笔者在分析散射变换原理和卷积神经网络(CNN)的基础上,通过低通滤波器作散射系数处理,结合Gram矩阵法来降低绝缘子串背景信息的噪声干扰,以增强低频系数的边缘纹理特征,结合SSD网络框架实现了CNN对绝缘子串实时定位计算的高效性。实验结果表明:该方法在保证实时计算的前提下,与传统SSD网络框架相比,召回率和交并比分别提升了1. 04%和1. 38%。

In view of the difficulties in locating insulator strings in transmission line images by UAV inspection,this paper analyzes the principle of scattering transformation and convolution neural network( CNN),processes the scattering coefficient through low-pass filter,and combines Gram matrix method to reduce the noise interference of insulator string background information,so as to enhance the edge ripple of low-frequency coefficient. Combined with the SSD network framework,the efficiency of CNN in real-time positioning of insulator strings is achieved. Test results show that compared with the traditional SSD network framework,the proposed method can improve the recall rate and the delivery-union ratio by 1. 04% and 1. 38% respectively under the premise of guaranteeing real-time computation.