输电线路图像绝缘子串实时定位方法
首发时间:2018-04-11
摘要:针对输电线路图像中存在绝缘子串定位难点,本文利用了散射变换具有弹性形变稳定性以及CNN学习能力强的特点,提出了一种速度快、精度高的新方法。首先对图像做散射变换提取不同尺度、方向的系数来抑制光照等干扰,其次计算格拉姆矩阵来增强低频系数的边缘纹理特征,再根据训练集数据计算网络特征图尺寸,然后将散射系数通过改进的YOLOv2网络得到初步定位,最后利用集成学习得到结果。实验结果表明:该方法在保证实时计算的前提下,与原YOLOv2网络相比,召回率可提升8.6%。
关键词: 绝缘子; 定位检测; 散射变换; 神经网络; 集成学习
For information in English, please click here
Real-time Locate of Instulator String in Transmission Line Images
Abstract:Considering the difficulty of locating the insulator string on aerial image, in this paper, a novel method based on the deformation stability of scatter transform coefficients is proposed. In this work, scatter transform is applied to extract coefficients to suppress the interference of light. To enhance the edge texture features, the Gram matrix of low frequency coefficients is calculated. Then, the size of network feature map is calculated according to the training set. With scatter transform coefficients, the improved YOLOv2 network is trained to generate the initial location of the insulator string, which is proved to be effective in object dectection. Finally, the accurate location can be obtained by ensemble learning. The experimental results demonstrate that the proposed method can not only improve the recall rate by 8.6% compared with the original YOLOv2 network, but also guarantee the real-time response speed.
Keywords: insulator locating scattering cnn ensemble learning
引用
No.****
动态公开评议
共计0人参与
勘误表
输电线路图像绝缘子串实时定位方法
评论
全部评论0/1000