第17卷第4期 智能系统学报 Vol.17 No.4 2022年7月 CAAI Transactions on Intelligent Systems Jul.2022 D0:10.11992/tis.202110035 网络出版地址：https:/ns.cnki.net/kcms/detail/23.1538.TP.20220421.0817.002.html 卷积神经网络金相组织自动识别 王佳锐2，刘能锋2，曲鹏 (1.廊坊燕京职业技术学院机电工程系，河北廊坊065200；2.哈尔滨工业大学实验与创新实践教育中心，广 东深圳518055) 摘要：为了降低人工分辨金相组织图像类别的误差率，提高分辨效率，采用卷积神经网络模型对金相组织图 像进行自动辨识。对制备金相样块所得铁素体与马氏体两种金相组织图像进行分析，提出符合金相组织图像 分布特征的预处理方案。通过采用图像尺寸归一化、灰度值归一化以及高斯平滑处理等方法，对原始金相组织 图像进行预处理，建立金相组织图像数据集。针对建立的铁素体和马氏体金相组织图像数据集，提出了适合金 相组织图像辨识的改进模型，分别记为LeNet-MetStr模型、AlexNet-MetStr模型和VGGNet-MetStr模型。对3种 改进卷积神经网络进行模型训练及分析，结果表明VGGNet-MetStr模型对2种金相组织图像自动辨识具有更 高的准确度。 关键词：卷积神经网络；金相组织；图像处理；网络模型；自动辨识；LeNet神经网络；AlexNet神经网络： VGGNet神经网络 中图分类号：TG141,TP391.4,TP183文献标志码：A 文章编号：1673-4785(2022)04-0698-09 中文引用格式：王佳锐，刘能锋，曲鹏.卷积神经网络金相组织自动识别.智能系统学报，2022,17(4)：698-706. 英文引用格式：WANG Jiarui,.LIU Nengfeng,.QU Peng.Automatic identification of metallographic structure based on convolution- al neural network J.CAAI transactions on intelligent systems,2022,17(4):698-706. Automatic identification of metallographic structure based on convolutional neural network WANG Jiarui,LIU Nengfeng QU Peng' (1.Mechanical and Electronic Engineering Department,Langfang Yanjing Polytechnic Inst.,Langfang 065200,China;2.Education Center of Experiments and Innovations,Harbin Institute of Technology,Shenzhen 518055,China) Abstract:The convolutional neural model was used to automatically identify metallographic structure images to reduce the error rate of manual resolution of metallographic structure image categories and improve the resolution effi- ciency.Two kinds of metallographic structure images of ferrite and martensite obtained from metallographic sample blocks were analyzed,and a preprocessing scheme conforming to the distribution characteristics of the metallographic structure image was proposed.Image size normalization,gray value normalization,and Gaussian smoothing are used to establish the metallographic image sample set and training set.Aiming at the established image data sets of two types of metallographic structures such as ferrite and martensite,the improved models suitable for metallographic structure image recognition are proposed,which are named the LeNet-MetStr model,AlexNet-MetStr model,and VGGNet-Met- Str model,respectively.Three improved convolutional neural networks were trained and analyzed.The results show that the VGGNet-MetStr model has higher accuracy for the automatic identification of two kinds of metallographic structure images. Keywords:convolutional neural network;metallographic structure;image processing;network model;automatic identi- fication:LeNet neural network:AlexNet neural network:VGGNet neural network 收稿日期：2021-10-29.网络出版日期：2022-04-21 钢铁材料是目前工业中应用最广、用量最大 基金项目：国家自然科学基金项目(52161004)：2021年廊坊市 的金属材料。钢铁材料受加热条件、轧制工艺、 科技局高新技术项目(2021011018). 通信作者：曲鹏.E-mail:372292920@qq.com 冷却速度、热处理工艺等因素的影响，其显微组DOI: 10.11992/tis.202110035 网络出版地址: https://kns.cnki.net/kcms/detail/23.1538.TP.20220421.0817.002.html 卷积神经网络金相组织自动识别 王佳锐1,2，刘能锋2 ，曲鹏1 （1. 廊坊燕京职业技术学院 机电工程系，河北 廊坊 065200; 2. 哈尔滨工业大学 实验与创新实践教育中心，广 东 深圳 518055） 摘 要：为了降低人工分辨金相组织图像类别的误差率，提高分辨效率，采用卷积神经网络模型对金相组织图 像进行自动辨识。对制备金相样块所得铁素体与马氏体两种金相组织图像进行分析，提出符合金相组织图像 分布特征的预处理方案。通过采用图像尺寸归一化、灰度值归一化以及高斯平滑处理等方法，对原始金相组织 图像进行预处理，建立金相组织图像数据集。针对建立的铁素体和马氏体金相组织图像数据集，提出了适合金 相组织图像辨识的改进模型，分别记为 LeNet-MetStr 模型、AlexNet-MetStr 模型和 VGGNet-MetStr 模型。对 3 种 改进卷积神经网络进行模型训练及分析，结果表明 VGGNet-MetStr 模型对 2 种金相组织图像自动辨识具有更 高的准确度。 关键词：卷积神经网络；金相组织；图像处理；网络模型；自动辨识；LeNet 神经网络；AlexNet 神经网络； VGGNet 神经网络 中图分类号：TG141;TP391.4;TP183 文献标志码：A 文章编号：1673−4785(2022)04−0698−09 中文引用格式：王佳锐, 刘能锋, 曲鹏. 卷积神经网络金相组织自动识别 [J]. 智能系统学报, 2022, 17(4): 698–706. 英文引用格式：WANG Jiarui, LIU Nengfeng, QU Peng. Automatic identification of metallographic structure based on convolution￾al neural network[J]. CAAI transactions on intelligent systems, 2022, 17(4): 698–706. Automatic identification of metallographic structure based on convolutional neural network WANG Jiarui1,2 ，LIU Nengfeng2 ，QU Peng1 (1. Mechanical and Electronic Engineering Department, Langfang Yanjing Polytechnic Inst., Langfang 065200, China; 2. Education Center of Experiments and Innovations, Harbin Institute of Technology, Shenzhen 518055, China) Abstract: The convolutional neural model was used to automatically identify metallographic structure images to reduce the error rate of manual resolution of metallographic structure image categories and improve the resolution effi￾ciency. Two kinds of metallographic structure images of ferrite and martensite obtained from metallographic sample blocks were analyzed, and a preprocessing scheme conforming to the distribution characteristics of the metallographic structure image was proposed. Image size normalization, gray value normalization, and Gaussian smoothing are used to establish the metallographic image sample set and training set. Aiming at the established image data sets of two types of metallographic structures such as ferrite and martensite, the improved models suitable for metallographic structure image recognition are proposed, which are named the LeNet-MetStr model, AlexNet-MetStr model, and VGGNet-Met￾Str model, respectively. Three improved convolutional neural networks were trained and analyzed. The results show that the VGGNet-MetStr model has higher accuracy for the automatic identification of two kinds of metallographic structure images. Keywords: convolutional neural network; metallographic structure; image processing; network model; automatic identi￾fication; LeNet neural network; AlexNet neural network; VGGNet neural network 钢铁材料是目前工业中应用最广、用量最大 的金属材料[1]。钢铁材料受加热条件、轧制工艺、 冷却速度、热处理工艺等因素的影响，其显微组 收稿日期：2021−10−29. 网络出版日期：2022−04−21. 基金项目：国家自然科学基金项目（52161004）；2021 年廊坊市 科技局高新技术项目（2021011018）. 通信作者：曲鹏. E-mail：372292920@qq.com. 第 17 卷第 4 期 智 能 系 统 学 报 Vol.17 No.4 2022 年 7 月 CAAI Transactions on Intelligent Systems Jul. 2022