摘要
无人机巡检航拍图像的品质对图像分类结果有很大的影响,为了准确分类航拍巡检图像,将包含绝缘子的图像分拣出来,需要大量高品质的图像样本。而航拍图像背景复杂、干扰多,缺乏高品质图像样本。为解决上述问题,采用3DS MAX软件进行绝缘子建模和渲染,生成绝缘子模拟图像,对样本库进行扩充。但是过量加入绝缘子模拟图像,会导致过学习,需要对加入的模拟图像数量进行限制。采用卷积神经网络算法,对扩充后的样本库进行训练,获取分类模型,用于分拣含绝缘子的图像。在此过程中,获取绝缘子真实图像样本数与模拟图像样本数的最优比例。实验结果表明,采用含有最优比例的图像样本库训练出的分类模型,能够对含绝缘子的图像进行准确分拣,分类准确率得到了大幅度提升。
The quality of Unmanned Aerial Vehicle （UAV） inspection images has great influence on image classification results. In order to classify the aerial inspection images accurately, and sort out the images contained the insulator, a large number of high - quality image samples are required. Due to complex background and a variety of interference for aerial images, massive high - quality image samples are needed. To solve this problem, sample set was expanded using the simulation insulator images which was modeled and rendered with 3DS MAX software. However, excessive addition of insulator simulation images may lead to overlearning, and the number of added simulation images is limited. The convolutional neural network algorithm was used to train the expanded sample set and to obtain a classification model for sorting out images contained insulators. During this process, the optimal proportion of the number of true image samples to the simulation ones was obtained. The experimental results show that using the classification model trained by the image sample set with the optimal ratio can sort out the images contained insulators accurately, and the classification accuracy has been greatly improved.
作者
程海燕
韩璞
董泽
张木柳
CHENG Hai -yan;HAN Pu;DONG Ze;ZHANG Mu -liu(Hebei Engineering Research Center of Simulation ＆ Optimized Control for Power Generation,North China Electric Power University,Baoding Hebei 071003,China)
出处
《计算机仿真》
北大核心
2018年第8期424-428,共5页
Computer Simulation
基金
河北省自然科学基金(F2017502016)
中央高校基本科研业务费专项资金项目(2014MS140)
关键词
图像分类
卷积神经网络
模拟图像样本
绝缘子
Image classification
Convolutional neural network
Simulation image samples
Insulator
