基于深度学习钢中非金属夹杂物图像识别  

Image recognition of non-metallic inclusions based on deep learning

作　　者：梁印 朱航宇[1,3] 罗林根 刘志 王宝[3] LIANG Yin;ZHU Hangyu;LUO Lingen;LIU Zhi;WANG Bao(Key Laboratory for Ferrous Metallurgy and Resources Utilization of Ministry of Education,Wuhan University of Science and Technology,Wuhan 430081,Hubei,China;Resource Application and Alloy Materials Division,China Iron and Steel Research Institute Group,Beijing 100081,China;Hubei Provincial Key Laboratory for New Processes of Ironmaking and Steelmaking,Wuhan University of Science and Technology,Wuhan 430081,Hubei,China)

机构地区：[1]武汉科技大学钢铁冶金及资源利用省部共建教育部重点实验室,湖北武汉430081 [2]中国钢研科技集团有限公司资源应用与合金材料事业部,北京100081 [3]武汉科技大学钢铁冶金新工艺湖北省重点实验室,湖北武汉430081

出　　处：《钢铁》2023年第12期62-70,共9页Iron and Steel

基　　金：国家自然科学基金资助项目(52074199,52374341)。

摘　　要：炼钢过程中钢水易和炉渣、耐火材料、气氛等相互作用形成非金属夹杂物,非金属夹杂物会破坏钢基体的连续性,增加钢组织的不均匀性,进而影响钢铁材料的塑性、韧性、抗疲劳强度等力学性能,成形过程中也易引起产品缺陷。夹杂物的定性检测一般通过扫描电子显微电镜(SEM)和能谱仪(EDS),但耗时较长、随机性也较大。因此,非金属夹杂物的快速检测和识别对改进炼钢工艺至关重要。近年来,随着计算机视觉技术的日渐成熟,基于区域的卷积神经网络(RCNN,region-based convolutional neural network)算法经过多代演化,并添加了掩码分支网络,形成了Mask-RCNN。Mask-RCNN既能实现夹杂物边框的准确定位,也能实现图像分割和识别分类,可有效应用于夹杂物分割和识别。采用计算机视觉(CV)任务中Mask-RCNN目标检测算法,对低密度钢中典型AlN、Al2O3、MnS和AlN-MnS 4类非金属夹杂物的SEM图片进行训练,经过10 000次的迭代训练后,对各类型夹杂物进行边框定位、图像分割及识别分类,并对测试集进行验证,实现了4类夹杂物边框的准确定位和图像分割。所选用模型对夹杂物检测识别效果较好,准确率高,4类目标夹杂物中,MnS和AlN-MnS夹杂物识别准确率达到100%,AlN夹杂物的识别准确率为95.91%,Al2O3夹杂物的识别准确率为83.33%。Due to the reaction between molten steel and slag,refractory,atmosphere,the non-metallic inclusions(NMIs)were generally formed during the steelmaking process.The NMIs may damage the continuity and increase the non-uniformity of the steel matrix,and then decrease the ductility,toughness,and fatigue resistance.The qualitative detection of NMIs was generally through SEM and EDS.However,the method was very time-consuming and had subjective and random.Rapid detection and identification of NMIs were very important to improve the steelmaking process.In recent years,with the increasing maturity of computer vision technology,RCNN target detection algorithms have evolved and improved for many generations,and the Mask-RCNN target detection algorithm was formed.Mask-RCNN can not only accurately realize the location of non-metallic inclusion,but also realize the mask segmentation and recognition classification.The Mask-RCNN can be effectively applied to inclusion segmentation and recognition.Based on the Mask-RCNN target detection algorithm,a large number of SEM images for AlN、Al2O3、MnS,and AlN-MnS NMIs in low-density steel were trained,and after 10000 iterations of training,the accurate location and mask segmentation of the mentioned four kinds of NMI were accurately identified.The model was effective and accurate for inclusion detection and recognition.The recognition accuracy of MnS and AlN-MnS inclusions was 100%,the value of AlN inclusion was 95.91%,and the value of Al2O3 inclusion was only 83.33%.

关 键 词：非金属夹杂物 图像识别 夹杂物分类 机器学习 Mask-RCNN 目标检测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术] TF703[冶金工程—钢铁冶金]