Abstract and subjects
Morphological features have proven to be a useful signature in determining the process histories of uranium oxides. Historically, morphological analysis has relied on using image analysis software to segment fully visible particles in Scanning Electron Microscopy (SEM) images of a particular sample and then compute attributes such as circularity, area, perimeter, ellipse aspect ratio of these segmented particles. One such software is Morphological Analysis for MAterial (MAMA) developed by Los Alamos National Laboratory. MAMA provides both segmentation and quantification functionality. Unfortunately, SEM images of nuclear materials can be difficult to segment due to overlapping particles, charging effects, and image clarity. It requires significant user inputs, which is time-consuming and tedious, to segment only fully visible particles. In this study, an alternative segmentation method, using a deep learning model, was used to segment fully visible particles. The deep learning model used in this study is a modified version of a well-known segmentation model in the computer vision community referred to as U-net. This model was able to produce the segmentation results similar to manual segmentation results obtained using MAMA with at least 85% accuracy in intersection over union metric. Furthermore, the model achieved a similar statistical relevance as manual segmentation under Kolmogorov-Smirnov (K-S) test.
[Display omitted]
•Segmentation of fully visible particles using a deep learning model.•Similar segmentation results between the proposed model and manual segmentation.•Differentiating materials processing histories using morphological features of the automatically segmented particles.