第4期 王佳锐，等：卷积神经网络金相组织自动识别 ·705· 点，采用图像增强、图像裁剪、尺寸缩放等方法进 材料进展，2020,39(5)：385-390 行图像预处理，得到能够反映铁素体和马氏体金 CAO Zhuo,DAN Yabo,LI Xiang,et al.Research on op- 相组织特征的训练样本集和测试集。 timization and prediction mechanism of material proper- 2)提出金相组织自动识别卷积网铬模型。结 ties based on gradient and feature analysis in convolution 合金相组织训练样本图像特征，优化卷积神经网 neural network[J].Materials China,2020,39(5):385- 络模型，对各层网络结构和参数进行调整，获得 390. [8]李维刚，湛竞成，范丽霞，等.基于卷积神经网铬的钢铁 训l练模型准确度优异的LeNet-MetStr、AlexNet- 材料微观组织自动辨识).钢铁研究学报，2020,32(1) MetStr、VGG-MetStr3种改进网络模型， 33-43. 3)通过对比3种改进卷积神经网络模型，分 LI Weigang,CHEN Jingcheng,FAN Lixia,et al.Auto- 析其性能，得出VGG-MetStr网络模型优于LeNet- matic identification of microstructure of iron and steel MetStr、AlexNet-.MetStr网络模型的结论，与理论 material based on convolutional neural network[].Journ- 相符。本文验证了使用卷积神经网络实现金相 al of iron and steel research,2020,32(1):33-43. 组织图像自动识别的可行性和准确性，为今后更 [9] 雷涛，李云形，周文政，等.数据与模型联合驱动的陶瓷 多金相组织种类的自动识别奠定基础，笔者今后 材料晶粒分割).自动化学报，2022,48(4)：1137-1152， 将对双相金相组织图像分析及自动辨识做进一步 LEI Tao,LI Yuntong,ZHOU Wenzheng,et al.Grain seg- 研究。 mentation of ceramic materials using data-driven jointing model-driven[J].Acta automatica sinica,2022,48(4): 参考文献： 1137-1152 [1]KHEDKAR P,MOTAGI R,MAHAJAN P,et al.A re- [10]GOLA J.BRITZ D.STAUDT T.et al.Advanced micro- view on advance higu strength steels[J].International structure classification by data mining methods[J].Com- journal of current engineering and technology,2016,6: putational materials science,2018,148:324-335 240. [11]WEBEL J,GOLA J,BRITZ D,et al.A new analysis ap- [2]HONEYCOMBE R W K.Steels:microstructure and proach based on Haralick texture features for the charac- properties[M].Fourth edition,Oxford:Butterworth- terization of microstructure on the example of low-alloy Heinemann,2017. steels[J].Materials characterization,2018,144:584- [3]KRAUSS G.Steels:processing,structure,and perform- 596. ance[M].2nd edition.Materials Park:ASM International. [12]TSUTSUI K.TERASAKI H.MAEMURA T,et al.Mi- 2015. crostructural diagram for steel based on crystallography [4]OHSER J,MUCKLICH F.Statistical analysis of micro- with machine learning[J].Computational materials sci- structures in materials science[].Practical metallography, ence,2019,159:403-411 2001,38(9):538-539. [13]GOLA J.WEBEL J,BRITZ D,et al.Objective micro- [5]张勇，李恒灿，梁明亮.基于PSO-BP神经网络的汽车 structure classification by support vector machine 用铸造AZ91镁合金晶粒尺寸的预测[.热加工工艺 (SVM)using a combination of morphological paramet- 2019,48(3):105-107,111. ers and textural features for low carbon steels[J].Com- ZHANG Yong,LI Hengcan,LIANG Mingliang.Predic- putational materials science,2019,160:186-196. tion of grain size of cast AZ91 magnesium alloy for auto- [14邹耀斌，雷帮军，臧兆祥，等.归一化互信息量最大化 mobile based on PSO-BP neural network[J].Hot working 导向的自动阈值选择方法[.自动化学报，2019， technology,2019,48(3:105-107,111 45(7):1373-1385 [6]张鹏，李靖，王文先，等.基于卷积神经网络模型的 ZOU Yaobin,LEI Bangjun,ZANG Zhaoxiang,et al. Gd2O,/6061A1中子屏蔽材料的力学性能预测U.原子 Automatic threshold selection guided by maximizing 能科学技术，2020,54(8)：1513-1518 normalized mutual information[J].Acta automatica sin- ZHANG Peng,LI Jing,WANG Wenxian,et al.Predic- ica,2019,45(7):1373-1385. tion of mechanical property of Gd2O3/6061Al neutron [15]WU Tingfan,MOVELLAN J.Semi-parametric Gaussi- shielding material based on convolutional neural network an process for robot system identification[C]//2012 model[]].Atomic energy science and technology,2020, IEEE/RSJ International Conference on Intelligent Ro- 54(8):1513-1518. bots and Systems.Vilamoura Algarve:IEEE,2012: [7]曹卓，但雅波，李想，等.基于卷积神经网络模型中梯度 725-731. 与特征分析的材料性能优化与预测机理研究[).中国 [16]VAN DER WILK M.RASMUSSEN C E.HENSMAN点，采用图像增强、图像裁剪、尺寸缩放等方法进 行图像预处理，得到能够反映铁素体和马氏体金 相组织特征的训练样本集和测试集。 2）提出金相组织自动识别卷积网络模型。结 合金相组织训练样本图像特征，优化卷积神经网 络模型，对各层网络结构和参数进行调整，获得 训练模型准确度优异的 LeNet-MetStr、AlexNet￾MetStr、VGG-MetStr 3 种改进网络模型。 3）通过对比 3 种改进卷积神经网络模型，分 析其性能，得出 VGG-MetStr 网络模型优于 LeNet￾MetStr、AlexNet-MetStr 网络模型的结论，与理论 相符。本文验证了使用卷积神经网络实现金相 组织图像自动识别的可行性和准确性，为今后更 多金相组织种类的自动识别奠定基础，笔者今后 将对双相金相组织图像分析及自动辨识做进一步 研究。 参考文献： KHEDKAR P, MOTAGI R, MAHAJAN P, et al. A re￾view on advance higu strength steels[J]. International journal of current engineering and technology, 2016, 6: 240. [1] HONEYCOMBE R W K. Steels: microstructure and properties[M]. Fourth edition, Oxford: Butterworth￾Heinemann, 2017. [2] KRAUSS G. Steels: processing, structure, and perform￾ance[M]. 2nd edition. Materials Park: ASM International, 2015. [3] OHSER J, MÜCKLICH F. Statistical analysis of micro￾structures in materials science[J]. Practical metallography, 2001, 38(9): 538–539. [4] 张勇, 李恒灿, 梁明亮. 基于 PSO-BP 神经网络的汽车 用铸造 AZ91 镁合金晶粒尺寸的预测 [J]. 热加工工艺, 2019, 48(3): 105–107,111. ZHANG Yong, LI Hengcan, LIANG Mingliang. Predic￾tion of grain size of cast AZ91 magnesium alloy for auto￾mobile based on PSO-BP neural network[J]. Hot working technology, 2019, 48(3): 105–107,111. [5] 张鹏, 李靖, 王文先, 等. 基于卷积神经网络模型的 Gd2O3 /6061Al 中子屏蔽材料的力学性能预测 [J]. 原子 能科学技术, 2020, 54(8): 1513–1518. ZHANG Peng, LI Jing, WANG Wenxian, et al. Predic￾tion of mechanical property of Gd2O3 /6061Al neutron shielding material based on convolutional neural network model[J]. Atomic energy science and technology, 2020, 54(8): 1513–1518. [6] 曹卓, 但雅波, 李想, 等. 基于卷积神经网络模型中梯度 与特征分析的材料性能优化与预测机理研究 [J]. 中国 [7] 材料进展, 2020, 39(5): 385–390. CAO Zhuo, DAN Yabo, LI Xiang, et al. Research on op￾timization and prediction mechanism of material proper￾ties based on gradient and feature analysis in convolution neural network[J]. Materials China, 2020, 39(5): 385– 390. 李维刚, 谌竟成, 范丽霞, 等. 基于卷积神经网络的钢铁 材料微观组织自动辨识 [J]. 钢铁研究学报, 2020, 32(1): 33–43. LI Weigang, CHEN Jingcheng, FAN Lixia, et al. Auto￾matic identification of microstructure of iron and steel material based on convolutional neural network[J]. Journ￾al of iron and steel research, 2020, 32(1): 33–43. [8] 雷涛, 李云彤, 周文政, 等. 数据与模型联合驱动的陶瓷 材料晶粒分割 [J]. 自动化学报, 2022, 48(4): 1137–1152. LEI Tao, LI Yuntong, ZHOU Wenzheng, et al. Grain seg￾mentation of ceramic materials using data-driven jointing model-driven[J]. Acta automatica sinica, 2022, 48(4): 1137–1152. [9] GOLA J, BRITZ D, STAUDT T, et al. Advanced micro￾structure classification by data mining methods[J]. Com￾putational materials science, 2018, 148: 324–335. [10] WEBEL J, GOLA J, BRITZ D, et al. A new analysis ap￾proach based on Haralick texture features for the charac￾terization of microstructure on the example of low-alloy steels[J]. Materials characterization, 2018, 144: 584– 596. [11] TSUTSUI K, TERASAKI H, MAEMURA T, et al. Mi￾crostructural diagram for steel based on crystallography with machine learning[J]. Computational materials sci￾ence, 2019, 159: 403–411. [12] GOLA J, WEBEL J, BRITZ D, et al. Objective micro￾structure classification by support vector machine (SVM) using a combination of morphological paramet￾ers and textural features for low carbon steels[J]. Com￾putational materials science, 2019, 160: 186–196. [13] 邹耀斌, 雷帮军, 臧兆祥, 等. 归一化互信息量最大化 导向的自动阈值选择方法 [J]. 自动化学报, 2019, 45(7): 1373–1385. ZOU Yaobin, LEI Bangjun, ZANG Zhaoxiang, et al. Automatic threshold selection guided by maximizing normalized mutual information[J]. Acta automatica sin￾ica, 2019, 45(7): 1373–1385. [14] WU Tingfan, MOVELLAN J. Semi-parametric Gaussi￾an process for robot system identification[C]//2012 IEEE/RSJ International Conference on Intelligent Ro￾bots and Systems. Vilamoura Algarve: IEEE, 2012: 725−731. [15] [16] VAN DER WILK M, RASMUSSEN C E, HENSMAN 第 4 期 王佳锐，等：卷积神经网络金相组织自动识别 ·705·