基于深度学习和数字图像处理的晶界分割与修复

Grain boundary segmentation and restoration based on deep learning anddigital image processing

通过综合应用深度学习与图像处理技术,对金相图像中晶界进行分割与修复,为晶粒度的准确评级提供基础。首先,通过采集并人工标注多样化的金相图像样本,利用U-net模型结合数据增强和合理的切割策略,提升模型的泛化能力和鲁棒性。其次,采用封闭晶界的数据集,增强模型的预测能力。第三,提出晶界匹配像素准确率这一新的评估指标,有效衡量了模型对晶界预测的准确性。此外,采用图像处理技术对提取到的晶界进行修复,进一步提高晶界线的完整性和连续性。这些方法的综合应用不仅提高了金相图像晶界分割的准确度,而且为材料微观结构的表征与性能评估提供了一种高效的技术手段。

By the aid of comprehensive application of deep learning and image processing technology,the grain boundaries in metallographic images are segmented and repaired,which provides a basis for accurate grain size rating.Firstly,by collecting and manually annotating diversified metallographic image samples,the U-net models combined with data enhancement and reasonable segmentation strategies are used to improve the generalization and robustness of the models.Secondly,a closed grain boundary data set is used to enhance the predictive ability of the models.Thirdly,a new evaluation index of pixel accuracy of grain boundary matching is proposed,which can effectively measure the accuracy of grain boundary prediction by the models.In addition,image processing techniques are applied to repair the extracted grain boundaries,which therefore can improve the integrity and continuity of the grain boundary lines.The comprehensive application of these methods not only improves the accuracy of grain boundary segmentation of metallographic images,but also provides an efficient technical means for the characterization and performance evaluation of material microstructures.