摘要
由霍尔-佩奇关系可知,组织的晶粒度直接影响纯铁材料的力学性能、耐腐蚀性和磁学特性,其标准化精确测量对纯铁材料性能评估至关重要。针对传统手工测量存在的效率瓶颈与操作者依赖性,以及现有图像处理方法和机器学习算法在晶界断裂修复和伪影抑制方面的不足,提出级联双阶段晶界分析框架。通过构建包含960幅纯铁金相图像的数据集,建立"识别-重建"协同机制。第1阶段采用改进U-Net(U-shaped convolutional network)架构实现晶界定位,结合加权损失函数解决晶界-晶粒比例失衡问题,保持97.3%晶界识别精度。第2阶段通过缺陷数据集训练重建网络,从而有效修复断裂晶界,使晶界完整度提升至92.7%。试验结果表明,本框架在关键指标上超越了对比方法,晶界识别的Dice系数达0.621,重建阶段的晶界闭环率提升至98.5%,单图处理耗时仅需0.28 s。在晶粒度测量方面,57.5%测试样本的相对误差低于5%,E_(MAP)(平均绝对百分比误差)为4.78%,在提升效率与客观性的同时,与人工测量结果高度一致。消融试验证实晶界重建与孪晶合并的协同作用使晶粒度测量EMA(平均绝对误差)降低69.7%,动态加权损失函数使IoU(交并比)指标提升了9.7%。该方法在深度学习框架中实现了晶界拓扑完整性与测量精度的协同优化,通过消除操作者依赖性提升了测量可重复性,为工业金相分析提供了标准化、可解释的自动化解决方案,有助于推进材料微观结构定量表征技术的发展。
Based on the Hall-Petch relationship,grain size directly influences the mechanical properties,corrosion resistance,and magnetic characteristics of pure iron materials,making its standardized and accurate measurement critical for material performance evaluation.To address the limitations of traditional manual methods,such as inefficiency and operator dependency,and the shortcomings of existing image processing and machine learning algorithms in repairing fractured grain boundaries and suppressing artifacts,a cascaded dual-stage grain boundary analysis frame⁃work was proposed.By constructing dataset of 960 pure iron metallographic images,"recognition-reconstruction"collaborative mechanism was established.In the first stage,enhanced U-Net(U-shaped convolutional network)architecture with weighted loss functions was employed to achieve grain boundary localization,addressing the class imbalance between grain boundaries and grains while maintaining 97.3%of recognition accuracy.The second stage utilized a defect-augmented dataset to train reconstruction network,effectively repairing fractured boundaries and improving grain boundary integrity to 92.7%.Experimental results demonstrate that the framework outperforms existing methods in key metrics,Dice coefficient of 0.621 for boundary recognition,98.5%closed-loop rate for reconstructed boundaries,and processing time of 0.28 s per image.For grain size measurement,57.5%of test samples exhibit relative errors below 5%,with a mean absolute percentage error(E_(MAP))of 4.78%,achieving high consistency with manual measurements while enhancing efficiency and objectivity.Ablation studies confirm that the synergy between boundary reconstruction and twin crystal merging reduces the mean absolute error(EMA)of grain size measurement by 69.7%,while the dynamic weighted loss function improves the Intersection-over-Union(IoU)by 9.7%.This approach achieves co-optimization of topological integrity and measurement accuracy within a deep learning framework.By eliminating operator dependency,the measurement repeatability has been enhanced,providing a standardized and interpretable automated solution for industrial metallographic analysis,which is conducive to promoting the development of quantitative characterization techniques for material microstructure.
作者
高志昊
GAO Zhihao(School of Jiluan Academy,Nanchang University,Nanchang 330031,Jiangxi,China)
出处
《钢铁》
北大核心
2025年第9期167-174,共8页
Iron and Steel
关键词
纯铁金相图像
深度学习
晶界识别
晶界重建
自动晶粒度测量
图像分割
缺陷修复
微观结构表征
pure iron metallographic image
deep learning
grain boundary detection
grain boundary reconstruction
automated grain size analysis
image segmentation
defect repair
microstructure characterization
