基于DenseNet和ResNet融合的发动机孔探图像分类研究

敖良忠; 马瑞阳; 杨学文

基于DenseNet和ResNet融合的发动机孔探图像分类研究

引用本文：敖良忠，马瑞阳，杨学文.基于DenseNet和ResNet融合的发动机孔探图像分类研究[J].计算技术与自动化,2021,(3):105-110

摘要点击次数: 463

全文下载次数: 0

作者	单位
敖良忠，马瑞阳，杨学文	（中国民用航空飞行学院航空工程学院,四川广汉 618307）

中文摘要:孔探是检测发动机内部损伤最重要的手段之一。为了解决发动机孔探检查中孔探人员主要依靠经验对损伤进行界定的问题，研究了基于DenseNet和ResNet融合的新型单通道网络结构，实现对发动机部件的分类，为后期孔探缺陷自动识别建立基础。通过对某大修厂孔探数据和自建数据进行处理，完成了孔探图像分类数据集的构建；训练新型的49层网络模型，在自建数据集测试集上测试的准确率和平均召回率分别为96.0%和95.9%，有较好的泛化能力，可以有效的对发动机部件进行分类。

中文关键词:发动机孔探部件分类 DenseNet ResNet 深度神经网络

Research on Engine Borescope Images Classification Based on Densenet and Resnet Fusion

Abstract:Borescope inspection is one of the most important means of detecting internal engine damage. In order to solve the problem that engineer mainly rely on experience to define damage during borescope inspection, a new single-channel network structure based on the fusion of DenseNet and ResNet was researched to realize the classification of engine components. At the same time, this establishes the foundation for the automatic identification of later flaw detection. By processing the borescope images data of a major repair plant and self-built data, the construction of the borescope image classification data set was completed. Trained a new type of 49-layer network model, with the accuracy of 96.0%, with the average recall rate of 95.9%, respectively, experimental results show that the network has good generalization ability and can effectively classify engine components.

keywords:engine borescope images parts classification Densenet Resnet deep neural network

查看全文 查看/发表评论 下载pdf阅读器