基于迁移学习的钢金相组织分类与识别方法的研究  

Research on Classification and Recognition Method of Steel Metallographic Structure Based on Transfer Learning

作　　者：张永志[1] 辛全忠 王永亮 孔祥明 刘昉 杨再胜 ZHANG Yongzhi;XIN Quanzhong;WANG Yongliang;KONG Xiangming;LIU Fang;YANG Zaisheng(College of Mechanical and Electrical Engineering,Inner Mongolia Agricultural University,Hohhot 010018,China;Inner Mongolia Energy Power Investment Group Co.,Ltd.,Electric Power Engineering and Technology Institute,Hohhot 010090,China)

机构地区：[1]内蒙古农业大学机电工程学院,呼和浩特010018 [2]内蒙古能源发电投资集团有限公司电力工程技术研究院,呼和浩特010090

出　　处：《材料导报》2021年第24期24152-24157,共6页Materials Reports

基　　金：国家自然科学基金(52061037);内蒙古农业大学高层次人才引进科研启动项目(NDYB2016-20)。

摘　　要：金相检验是分析钢内部组织的常用方法,其中检验图像由人工判别,容易受到主观因素的影响而造成结果的不确定。近年来,深度学习(Deep learning,DL)方法中的卷积神经网络(Convolutional neural networks,CNN)能从原始图像中学习复杂的特征,在图像分类与识别领域得到了广泛的应用。CNN建模需要大量的训练样本才能达到较好的泛化能力,材料科学与工程领域针对具体问题的数据集往往较小,不能满足CNN建模的条件,制约了DL在材料领域的应用。本研究基于ImageNet数据集预训练VGG19模型,对火力发电机组耐热钢金相检验图像进行识别,采用冻结全部卷积层权值和微调部分卷积层权值两种迁移学习方法,可以克服金相图像数据集较小的问题,实现小样本数据集的深度学习建模,两种方法的准确率分别为92.5%和94.2%。微调方式的迁移学习CNN模型具有较快的收敛速度、较高的训练精度与泛化能力,能够对金相组织图像进行较为准确的分类与识别,是一种智能的钢金相组织识别方法,也是自动化分类与识别钢金相组织的一种新方法。Metallographic examination is a common method to analyze the internal micro-structure of steel,in which the image discrimination is manually processed and is consequently prone to result uncertainty due to subjective factors.Recently,convolutional neural networks(CNN)in the deep learning(DL)method is widely used in the field of image classification and recognition for its capability of learning complex features from raw images.CNN modelling requires large training samples to make a good generalization ability while in material science and engineering,data set for specific actual problems are usually too small to make a good modelling and hence put limit to the DL application.In this study,VGG19 model was pre-trained based on ImageNet dataset to classify and recognize images of heat resisting steel for thermal-electric generator,using two transfer learning methods:freezing all convolutional layers’weights and fine-tuning partial convolutional layers’weights,which realized the deep learning modeling of small sample data sets,overcome the problem of small metallographic set,and identify the metallographic inspection images of heat-resistant steel of thermal power generating units,with the accuracy rates of 92.5%and 94.2%,respecfively.The transfer learning CNN model with fine-tuning has fast convergence speed,high training accuracy and generalization ability,which can accurately classify and identify metallographic images.It is a method for intelligently identifying the metallographic structure of steel,and also a new method for automatic classification and identification of steel metallographic organization.

关 键 词：深度学习 卷积神经网络 迁移学习 金相组织 分类与识别 

分 类 号：TG142.15[一般工业技术—材料科学与工程]