摘要
针对现有深度学习方法在复杂背景下无法保证裂纹特征映射的有效传递和融合,结构化森林对相似、随机的裂纹特征无法准确判别的问题,提出一种基于全卷积神经网络和结构化森林的结构体裂纹分割方法。首先,以全卷积神经网络框架为基础构建5种消融神经网络用于扩充细微裂纹全局特征;其次,提出一种基于多尺度结构化森林的裂纹分割参数竞争策略用于效改善细微裂纹区分能力;最后利用基于消融神经网络和结构化森林的耦合分割方法进行裂纹图像的联合预测。在2类结构体裂纹数据集上对所提方法进行实验验证表明,所提方法能够在复杂相似背景下提高裂纹检测的精度,可以实现有效的结构体健康监测。
Aiming at the problems that existing deep learning method cannot ensure effective transfer and fusion of crack feature mapping in complex background, and structured forest cannot accurately distinguish the crack features with similar behaviors and random characteristics. This paper proposes a structure crack segmentation method based on fully convolutional network and structured forests. Firstly, five ablation neural networks are constructed based on the framework of full convolution neural network to extend global characteristics of microcracks. Secondly, a competitive strategy of crack segmentation parameters based on multi-scale structured forests is proposed to effectively improve the ability to distinguish tiny crack. Finally, the coupling segmentation method based on ablation neural network and structured forests is used to jointly predict the crack images. The experiments on two kinds of structural crack datasets were carried out to verify the proposed method, experiment results illustrate that the proposed method can improve the crack detection accuracy under complex and similar backgrounds, and can achieve effective structural health monitoring.
作者
王森
伍星
张印辉
柳小勤
Wang Sen;Wu Xing;Zhang Yinhui;Liu Xiaoqin(School of Mechanical and Electrical Engineering,Kunming University of Science and Technology,Kunming 650500,China;Yunnan Vocational College of Mechanical and Electrical Technology,Kunming 650203,China)
出处
《仪器仪表学报》
EI
CAS
CSCD
北大核心
2020年第8期170-179,共10页
Chinese Journal of Scientific Instrument
基金
国家自然科学基金(51875272,61761024)
云南省级人培项目(KKSY201801018)
云南省教育厅科学研究基金(2019J0045)项目资助
关键词
全卷积神经网络
结构化森林
反对称双正交小波变换
结构体裂纹分割
fully convolutional neural network
structured forest
anti-symmetrical biorthogonal wavelet transform
structure crack segmentation
作者简介
王森,2007年于郑州轻工业学院获得学士学位,分别在2014年和2017年于昆明理工大学获得硕士学位和博士学位,现为昆明理工大学讲师、硕士生导师,主要研究方向为机器视觉检测与图像识别。E-mail:wangsen0401@126.com;通信作者:伍星,分别在1994年和1997年于昆明理工大学获得学士和硕士学位,2005年于上海交通大学获得博士学位,现为云南机电职业技术学院教授,昆明理工大学教授、博士生导师,主要研究方向为信号处理及设备智能诊断。E-mail:xwu@kust.edu.cn
