TC4-DT合金中片状α相的高精度定量分析方法

High-precision quantitative analysis methods of lamellarαphase in TC4-DT alloy

针对网篮组织片状α相体积分数难以精确定量分析以及粘连α相难分离表征的问题,结合体视学原理,采用随机森林、遗传算法和改进遗传算法对TC4-DT合金网篮组织片状α相进行表征。首先,预处理采集网篮组织图像;然后,利用样本中片状α相和β相特征对随机森林模型进行训练。考虑到传统遗传算法图像分割易陷入局部最优解以及收敛速度过快的问题,本文采用精英选择和轮盘赌结合的方法初始化种群,设计了两段式交叉概率和抛物线型变异概率优化遗传算法。最后,利用Java程序验证随机森林模型并自动定量分析片状α相的体积分数,结合实例定量分析片状α相的特征参数。结果表明:采用改进遗传算法运行时时间缩短60%,且图像处理效果也得到提升;随机森林模型不仅在训练样本中的分类准确率达到99.89%,而且在测试样本中的准确率也达到99.29%。这说明随机森林模型能精确地分离片状α相与β相且具有较好的泛化能力。

It is difficult to quantify the volume fraction of lamellarαphase in basket-weave microstructure accurately and to characterize the adhesiveαphase by separation.In order to solve these problem,the random forest,genetic algorithm and improved genetic algorithm were used to characterize the lamellarαphase in basketweave microstructure of TC4-DT alloy combined with the principle of stereology.Firstly,the collected basketweave microstructure images were preprocessed.Then,the random forest model was trained by using the characteristics of lamellarαphase andβphase in the sample.Considering that the traditional genetic algorithm image segmentation are prone to fall into the local optimal solution and the convergence speed is too fast,the combination method of elite selection and roulette was used to initialize the population in the paper.Also,a twostage crossover probability and parabolic mutation probability optimization genetic algorithm was designed.Finally,the Java program was used to verify the random forest model and quantitative analysis the volume fraction of lamellarαphase automatically,and quantitative analysis of the lamellarαphase characteristic parameters were performed based on the examples.The results show that the run time of the improved genetic algorithm is shortened by 60%and the effect of image processing is improved.The classification accuracy of the random forest model not only reaches 99.89%in the training samples,but also reaches 99.29%in the test samples.It is proved that the lamellarαphase andβphase can be separated accurately by the model and its generalization performance is better.