基于迁移学习和Xception网络的混凝土裂纹识别研究

2022-07-21 11:54:35徐峥匀钱松荣
软件工程 2022年7期

徐峥匀 钱松荣

摘  要:针对深层卷积神经网络检测表面结构裂纹耗费时间长、精度不够高的问题,基于Xception网络进行自适应调整重构其分类器,利用图像增广技术扩充数据集后,引入迁移学习的方法对Xception网络进行训练。同时,与构建的ResNet50、InceptionV3和VGG19三个深层卷积神经网络模型进行对比实验,重新验证其性能。实验证明,引入迁移学习不仅可以提升模型的整体性能,还能缩减训练深层卷积神经网络的时间,训练的模型在数据集上的识别精确率达到96.24%,在对比实验中达到96.50%。

关键词:迁移学习;卷积神经网络;图像识别;图像增广

中图分类号:TP399     文献标识码:A

Research on Concrete Crack Identification based on Transfer Learning and Xception

XU Zhengyun1, QIAN Songrong1,2

(1.College of Mechanical Engineering, Guizhou University, Guiyang 550025, China;

2.State Key Laboratory of Public Big Data, Guizhou University, Guiyang 550025, China)

xu.zhengyun@qq.com; 909718747@qq.com

Abstract: Aiming at the problems of time consuming and low accuracy of deep convolution neural network in detecting surface structural cracks, this paper proposes to adaptively adjust and reconstruct its classifier based on Xception, and use image augmentation technology to expand the dataset. After that, transfer learning method is introduced for training Xception. At the same time, its performance is re-verified by comparing it with three deep convolution neural network models, ResNet50, InceptionV3 and VGG19. Experiments show that the introduction of transfer learning not only improves the overall performance of the model, but also reduces the time of training the deep convolutional neural network. Recognition accuracy of the trained model on the dataset reaches 96.24%, and 96.50% in the comparative experiment.

Keywords: transfer learning; convolutional neural network; image recognition; image augmentation

1   引言(Introduction)

桥梁结构会随着时间的推移而退化,从而出现裂纹甚至坍塌,造成安全事故[1]。深度学习[2](Deep Learning)技术处理大量数据和自动化特征提取的能力使其能够开发混凝土结构裂缝检测问题的解决方案。而迁移学习[3](Transfer Learning)可以从相关或者相近领域中转移已经标注好的数据、参数或者知识结构,从而完成或改进目标领域或目标任务的学习效果。

本文基于迁移学习和Xception[4]网络,使用混凝土裂纹图像数据集训练出能够检测裂纹的模型,利用图像增广技术扩充数据集;比较了使用扩增数据集和迁移学习前后模型在测试集上的表现,得到一个训练时间短并且具有较高精度的网络模型。

2   实验流程(Experimental process)

基于迁移学习的Xception网络裂纹识别模型建模流程如图1所示。首先,对原有数据集的图片进行图片增强,从而增加训练集;然后将……

登录APP查看全文