基于深度卷积网络的复杂环境中绝缘子串特征检测

Research on insulator string detection based on improved convolutional neural network

摘要:

绝缘子串工作状态的智能监测是实现智能电网的重要前提。针对电网巡检图像中绝缘子串特征检测效果受复杂环境的影响。本文引入超像素和统计方法来降低复杂场景的难度,首先采用多尺度方式将图像分解为不同的层,然后采用SLIC将每层图像分解为不同大小的超像素,采用超像素特征来描述多尺度图像的某个区域位置的特征信息,得到图像的粗略显著区域,并作为样本集输入区域网络Region net进行处理,最后通过网络训练迭代得到准确完整的显著特征图。将本文算法和其它流行算法对不同环境中绝缘子串图像进行特征检测对比实验,证明本文算法的F-Measure 以及平均误差MAE均优于当前流行算法。

Abstract:

The real time intelligent monitoring of the working state of the insulator string is an important prerequisite for the realization. The detection effect of insulator strings is affected by the environment in the current patrol image. This article introduced the super pixels and statistical methods to reduce the difficulty of complex scenes, using multi-scale approach first the image is decomposed into different layers, and then uses the SLIC will each layer of the image is decomposed into different sizes of super pixels, super characteristics of pixels is used to describe the multi-scale, the feature information of the image of a position, and comprehensive get images of rough area, and as a sample set input area network Region net for processing. Finally, accurate and complete salient features are obtained through network training iteration. By comparing the characteristics of insulator string images in different environments with other popular algorithms, it is proved that the F-measure and mean error MAE of this algorithm are superior to current popular algorithms.