基于深度学习算法的大尺寸铝合金中第二相的识别提取与定量统计分析

Identification and quantitative statistical analysis of second phase in aluminum alloy based on deep learning algorithm

摘要:

铝合金中第二相粒子是在铸造过程中产生并对材料的物理化学性能有着显著影响的显微组织。目前常用的第二相的定量表征方法存在人工工作量和消耗时间过大的问题。本文提出一种基于深度学习的大尺寸铝合金中第二相的快速提取与定量统计表征方法，通过对图像中的第二相特征的快速、智能化提取，实现多角度精细化的定量统计。研究结果显示，该方法的图像处理时间与软件批量处理时间一样为0.4s/张，但图像分割精度从42.74%提升到91.12%。从数据集制作方面，本方法用MIPAR软件结合人工微调代替传统的手工标记，大大节省了人工时间。为了满足精细化表征的要求，对11万张包括四种类型的全视场7B05铝合金扫描电镜图像进行分割提取，计算了第二相的长宽比、最短间距、面分布以及线分布等新的表征参数，其中线分布结果证明，与传统的随机选取视场的统计结果相比较，本文使用的全视场多角度的统计表征方式误差更小，特征信息更全面。定量统计结果显示，厚度为6mm的铸轧7B05-T4铝合金截面，在最近邻上下表面的位置第二相的平均面积最小，在厚度在3mm的位置存在波谷值；将截面均分为三个区域，第二相的平均面积分别为1.99μm2、1.84μm2、2.18μm2，平均长宽比分别为1.89、1.95、1.84，第二相数量分别为33574、33207、42035个。以上结果表明，基于深度学习的铝合金中第二相的快速提取与定量统计表征方法可进行多角度的分析数据与挖掘，为组织、性能研究提供依据。

Abstract:

The second phase particles in aluminum alloy are produced in the casting process and have a significant effect on the physical and chemical properties of the material. At present, the commonly used quantitative characterization methods of the second phase have the problems of too much manual work and time consumption. In this paper, a fast extraction and quantitative statistical characterization method of the second phase in large-scale aluminum alloy based on deep learning is proposed. This method can achieve multi angle refined quantitative statistics by fast and intelligent extraction of the second phase features in the image. The results show that the image processing time of this method is the same as that of software batch processing, which is only 0.4s/sheet, but the image segmentation accuracy is improved from 42.74% to 91.12%. In order to meet the requirements of fine characterization, 110000 full field sem images of 7B05 aluminum alloy, including four types, were segmented and extracted. The new characterization parameters of the second phase, such as aspect ratio, shortest distance, surface distribution and line distribution, were calculated. The results of line distribution show that the full field multi angle method used in this paper is better than the traditional statistical results of randomly selected field of view. This method shows that the error of statistical representation is smaller and the feature information is more comprehensive. The quantitative statistical results show that the average area of the second phase is the smallest at the nearest upper and lower surface of the cast rolled 7b05-t4 aluminum alloy section with a thickness of 6 mm, and there is a trough value near the thickness of 3 mm; The average area of the second phase is 1.99 μ m2, 1.84 μ m2 and 2.18 μ m2, the average aspect ratio is 1.89, 1.95 and 1.84, and the number of the second phase is 33574, 33207 and 42035, respectively. The above results show that the rapid extraction and quantitative statistical characterization method of the second phase in aluminum alloy based on deep learning can carry out multi angle data analysis and mining, and provide the basis for the study of microstructure and properties.