Extracting accurate volume fraction and size measurements of γ″ and γ′ precipitates in iron-based superalloys from micrographs is challenging and conventionally involves manual image processing due to their smaller size, and similar crystal structures and chemistries. The co-precipitation of composite particles further complicates automated segmentation. In this work, different types of traditional machine learning approaches and a convolutional neural network (CNN) were compared to a non-machine learning approach, for the segmentation of the composite particles of γ″ and γ′ precipitates. The objective was to optimize metrics of segmentation accuracy and the required computational resources. The data set contains 47 experimentally generated scanning electron micrographs of IN718 alloy samples, computationally increased to 188 images (900 × 900 px). All algorithms are containerized using singularity, publicly available, and can be modified without dependencies. The CNN and the random forest models achieve 95% and 94% accuracy, respectively, on the test images with better computational efficiency than the non-machine learning algorithm. The CNN tested accurately over a range of imaging conditions.

提取的准确体积分数和尺寸测量γ “和γ从显微照片中的铁基超耐热合金”析出物是富有挑战性和以往涉及手动图像处理，由于其较小的尺寸，以及类似的晶体结构和化学组成。复合颗粒的共沉淀进一步使自动分割变得复杂。在这项工作中，将不同类型的传统机器学习方法和卷积神经网络（CNN）与非机器学习方法进行了比较，以分割γ ″和γ的复合粒子沉淀。目的是优化分割精度和所需计算资源的度量。该数据集包含47个实验生成的IN718合金样品的扫描电子显微照片，计算得出的图像数量增加到188张（900×900 px）。所有算法都使用奇异性进行容器化，可公开获得，并且可以在没有依赖的情况下进行修改。CNN模型和随机森林模型在测试图像上的准确率分别达到95％和94％，与非机器学习算法相比，具有更高的计算效率。CNN在一系列成像条件下均进行了准确测试。