【免费】基于MaskR-CNN及Yolov4的电力巡检中绝缘子缺陷研究

需积分: 0 94 浏览量 2022-03-26 23:35:47 上传评论 1 收藏 3.83MB PDF 举报

资源详情

资源评论

资源推荐

基于 Mask R-CNN 及 Yolov4 的电力巡检中绝缘子缺陷研究

摘要

输电线路是国家电力系统中负责电能输送的一个重要的组成部分。架空输电

线路长期经历着电闪雷击、风吹日晒、污秽、雨雪及沉陷等外界环境的侵害。目

前传统的电力巡检仍是人工巡检，新型技术利用无人机搭载高分辨摄像机，近距

离拍摄线路及其关键部件获取航拍图像来代替人工巡检。其中，航拍图像中输电

线路部件的识别及缺陷检测是关键技术也是技术难点。本文提出一种利用 Mask

R-CNN 技术提取绝缘子串珠的掩膜图像、采用 Yolo-v3 和 Yolo-v4 进行对照，通

过神经网络模型对其绝缘子自爆区域进行特征提取，特征判断，识别定位。具体

流程如下：

针对问题一中绝缘子串珠分割，本文根据不同绝缘子类型进行图像类别划

分，通过图像分割优化模型对小目标进行数据集制作。采用 Mask R-CNN 为框架

构建了深度学习算法，并以算法为核心搭建了绝缘子串图像检测平台。最后选取

数据集 80%进行模型训练，20%进行测试，使用极大值抑制算法，得出最佳先验

框（best_anchor），对经过模型得出的绝缘子掩模图与官方的绝缘子掩模图进行

计算，求出评价模型的 Dice 系数为 0.83。

针对问题二中绝缘子自爆识别和定位，对 35 张原始含有自爆区域高分辨率

图像，经由图像分割、图像增强、归一化大小等对图像进行预处理，将得到的数

据集用 LabelImg 工具进行标记，将标记后的数据集制作为 VOC2007 格式，随机

选取 80%数据集进行模型训练，20%数据集放进 Yolo-v3 和 Yolo-v4 神经网络进行

模型测试，训练至 loss 函数收敛，并用 Mean AP（平均精度均值）计算得出正确

率为 94.7%，与 GroundTruth 绝缘子自爆区域进行对比得出该模型平均 IOU（重

叠度）为 82.5%。

关键词：电力巡检，缺陷检测，Mask R-CNN，掩膜，Yolo-v4，识别定位

Abstract

Transmission lines are an important part of electric power transmission in

national power system.The overhead transmission lines have long been damaged by

lightning, wind and sun, pollution, rain and snow, and subsidence.At present, the

traditional electric power inspection is still manual inspection. The new technology

USES unmanned aerial vehicles (uavs) equipped with high-resolution cameras to

capture aerial images of circuits and key components in close range to replace manual

inspection.Among them, the identification and defect detection of transmission line

components in aerial images are key technologies and technical difficulties.In this

paper, a Mask image of insulator beads is extracted by using Mask r-cnn technology,

compared with Yolo-v3 and Yolo-v4, and features of the self-exploding region of the

insulator are extracted, identified and located by using neural network model.The

specific process is as follows:

For the first question,insulator bead segmentation, the image classification is

carried out, and the data sets of small targets are made by the image segmentation

optimization model. A deep learning algorithm is constructed by using Mask R-CNN

as the frame, and an insulator string image detection platform is built with the

algorithm as the core. finally ,80% of the data set is selected for model training ,20%

is tested, and the best prior frame (best_anchor) is obtained by using the maximum

suppression algorithm. the insulator mask obtained by the model and the official

insulator mask are calculated. the Dice coefficient of the evaluation model is 0.83.

For the second question, insulator self-explosion recognition and localization,

first for 40 data original image, the image is preprocessed by image segmentation,

image enhancement, normalized size, etc. the obtained data set is labeled with

LabelImg tools, and the Yolo-v4 neural network is used to build the model. finally, the

labeled data set system is used as the VOC2007 format ,80% is randomly selected for

model training ,20% is trained to loss function convergence, and the correct rate is

94.7% calculated by using Mean AP (mean precision). compared with the

GroundTruth insulator self-explosion region, the average iou( overlap) of the model is

82.5%.

Keywords: Electric power inspection,Defect detection ,Mask R-CNN,mask,Yolo-v4,

identify location

 
目录 
一、 绪论........................................................... 1 
1 研究背景..................................................... 1 
2 国内研究现状................................................. 1 
3 研究目标及流程............................................... 1 
二、 数据预处理..................................................... 3 
1 图像分割..................................................... 3 
2 图像金字塔................................................... 4 
3 数据增强..................................................... 4 
4 归一化图像大小 .............................................. 5 
三、 绝缘子串珠图像分割............................................. 6 
1 Mask R-CNN 神经网络结构 .................................... 6 
2 训练样本的制作及标注 ........................................ 6 
3 Mask R-CNN 训练过程 ........................................ 8 
3.1 部分训练参数设置 ....................................... 8 
3.2 训练过程 ............................................... 9 
4 Mask R-CNN 训练结果分析 .................................... 9 
4.1 检测及分割过程......................................... 9 
四、 绝缘子自爆区域识别与定位...................................... 11 
1 数据集制作.................................................. 11 
2 样本标记及分类.............................................. 11 
3 基于 Yolo-v3 的自爆区域识别与定位 ............................ 13 
3.1 Yolo-v3 网络结构 ...................................... 13 
3.2 Yolo-v3 训练过程可视化分析 ............................ 14 
3.3 Yolo-v3 训练结果分析 .................................. 15 
4 基于 Yolo-v4 的自爆区域识别与定位 ............................ 17 
4.1 Yolo-v4 网络结构 ...................................... 17 
4.2 Yolo-v4 训练可视化分析 ................................ 18 
4.3 Yolo-v4 训练结果分析 .................................. 20 