基于深度学习的超低碳钢显微组织特征强化与精细化分析  

Enhancement and refinement of microstructure characteristics of ultra-low carbon steel based on deep learning

作　　者：王煜乔 吴思炜 曹光明[1] 刘建军[1] 窦君 闫新悦 刘振宇[1] WANG Yuqiao;WU Siwei;CAO Guangming;LIU Jianjun;DOU Jun;YAN Xinyue;LIU Zhenyu(State Key Laboratory of Rolling Technology and continuous Rolling Automation,Northeastern University,Shenyang 110819,Liaoning,China)

机构地区：[1]东北大学轧制技术与连轧自动化国家重点实验室,辽宁沈阳110819

出　　处：《钢铁》2025年第2期119-127,共9页Iron and Steel

基　　金：国家自然科学基金青年科学基金资助项目(52104370);兴辽英才资助计划资助项目(XLYC2203186);中国博士后科学基金资助项目(2022T150205);东北大学博士后基金资助项目(20210203);轧制技术及连轧自动化国家重点实验室自主课题资助项目(ZZ2021005);中信铌钢发展奖励基金资助项目(2022-M1824)。

摘　　要：超低碳钢显微组织为铁素体,在制样过程中极易出现划痕和晶界腐蚀不清晰的现象,严重影响金相组织分析。同时,显微组织特征的分析结果严重依赖于专家经验,受主观因素影响较大且效率低。为了高效获得超低碳钢显微组织特征信息,基于超低碳钢金相图像数据集,采用归一化、自适应阈值法处理图像,增强图像对比度;融合自注意力机制(Self-Attention,SA)和循环回归生成对抗神经网络(CycleGan),开发基于CycleGan+SA的晶界增强算法;建立超低碳钢显微组织特征强化模型,实现了显微组织图像的自动处理与晶界信息的特征强化。在此基础上,采用分水岭分割算法对晶界强化后的显微组织图像进行精细化分析。结果表明,CycleGan+SA算法可以有效去除原始金相图像中的划痕并补全晶界模糊区域,实现超低碳钢晶界特征强化。相比原始的CycleGan算法,引入注意力机制后,CycleGan+SA算法可以实现更清晰的晶粒分割,图像识别精确度P值由97.43%提升至98.75%,综合评价指标F值由97.49%提升至98.73%。在显微组织精细化分析方面,通过与常用分析软件对比,超低碳钢显微组织特征强化模型与Image J软件测定的晶粒尺寸平均误差为1.2个晶粒,与Image Pro Plus软件测定的晶界比例误差为0.008个百分点,模型与软件统计结果吻合较好,具备一定的应用前景。The microstructure of ultra-low carbon steel is ferrite,which is prone to scratches and unclear grain boundary corrosion in the process of sample preparation,it is seriously affects the metallographic analysis.At the same time,the analysis results of microstructure characteristics are heavily dependent on expert experience,subject to subjective factors and low efficiency.In order to obtain the microstructure characteristic information of ultra-low carbon steel efficiently,based on the ultra-low carbon steel metallographic image data set,the normalized and adaptive threshold method was used to process the image and enhance the image contrast.A grain boundary enhancement algorithm based on CycleGan+SA was developed by integrating Self-Attention(SA)and cyclic regressiongenerating adversarial neural network(CycleGan).The microstructure feature enhancement model of ultra-low carbon steel was established to realize the automatic processing of microstructure images and the feature enhancement of grain boundary information.On this basis,the watershed segmentation algorithm is used to analyze the microstructure image after grain boundary enhancement.The results show that the CycleGan+SA model can effectively remove the scratches in the original metallographic image and complete the fuzzy grain boundary region,so as to realize the grain boundary enhancement of ultra-low carbon steel.Compared with the original CycleGan algorithm,after the introduction of attention mechanism,CycleGan+SA algorithm can achieve clearer grain segmentation,and the image recognition accuracy is increased from 97.43%to 98.75%,and the F value is increased from 97.49%to 98.73%.In terms of microstructure refinement analysis,compared with commonly used analysis software,the average error of grain size measured by ultra-low carbon steel microstructure feature enhancement model and Image J software is 1.2 grains,and the error of grain boundary ratio measured ultra-low carbon steel microstructure feature enhancement model and Image Pro Plus software is 0.

关 键 词：CycleGan 自注意力机制 深度学习 晶界特征强化 超低碳钢 图像识别 分水岭分割算法 超低碳钢 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TG142.1[自动化与计算机技术—控制科学与工程]