基于新型深度神经网络的民机表面缺陷识别
投稿时间:2019-08-13  修订日期:2019-08-16  点此下载全文
引用本文:
摘要点击次数: 154
全文下载次数: 0
作者单位邮编
张德银 中国民用航空飞行学院 航空工程学院 四川 广汉 618307 618307
陈从翰* 中国民用航空飞行学院 航空工程学院 四川 广汉 618307 618307
黄选红 中国民用航空飞行学院 航空工程学院 四川 广汉 618307 
徐志强 中国民用航空飞行学院 航空工程学院 四川 广汉 618307 
基金项目:(J2019-088)
中文摘要:为解决机务人员依靠经验来对民航飞机的表面缺陷进行识别时易发生误判的问题,本文开发了一种用于民机表面的缺陷识别的结合Inception-net、残差模块的新型深度神经网络。首先,通过对各机场的在修飞机表面缺陷进行采样建立数据集,手段包括使用图像处理修复不合格图像、使用数据增强缓解数据类别不平衡、使用立方卷积插值法降采样保留图像特征等图像预处理操作。然后在自建的数据集上对新型深度神经网络与其他神经网络进行对比测试。最终实验结果表明,新型神经网络在较少的参数下能够达到最深的网络深度,且在自建数据集的测试集上的识别率和查全率分别为74.23%和62.29%,优于进行对比的其他网络。说明在一定程度上该网络能够有效用于民机表面缺陷识别工作中。
中文关键词:民航飞机  表面缺陷识别  残差  Inception-net  深度神经网络  
 
Investigation of Civil Aviation Aircraft Surface Defects Recognition Based on novel Designed Deep Neural Network
Abstract:A novel designed neural network combining the Inception_net and the residual module is employed to identify surface defects of civil aviation aircraft in order to solve the problems of that the maintenance crew are easy to misjudge surface defects of civil aviation aircraft.Firstly, the data set has been established by sampling surface defects of aircraft from various airlines and overhaul factory.Including the use of image processing methods to repair unqualified images; the use of data enhancement to alleviate the problem of data category imbalance; the use of cubic convolution interpolation to downsample and preserve image features and other operations to image preprocessing data.Secondly, the new designed neural network is compared with other neural networks on self-built data sets.The experimental results show that the new neural network has the deepest network depth with fewer parameters in all tested networks. And the recognition rate and recall rate on the test set are 74.23% and 62.29%, respectively, which is better than other neural networks.It shows that the network can be effectively used for civil aircraft surface defect identification to some extent.
keywords:civil aviation aircraft  surface defect recognition  residual  Inception-net  deep neural network
查看全文   查看/发表评论   下载pdf阅读器