| About this Abstract |
| Meeting |
MS&T23: Materials Science & Technology
|
| Symposium
|
2023 Undergraduate Student Poster Contest
|
| Presentation Title |
Detectron2 is a Machine Learning Object Detection Algorithm That Was Utilized for the Automated Detection of Fibers Within Composite Materials |
| Author(s) |
Diana Johnson, Craig Przybyla, Ashley Hilmas, Mathew Schey |
| On-Site Speaker (Planned) |
Diana Johnson |
| Abstract Scope |
Machine learning is being leveraged with great success to improve the speed and efficiency of segmentation and characterization of materials. Here we show how a Region-based Convolutional Neural Network (R-CNN) based model in the open-source package, Detectron2, has been trained at AFRL to identify fibers in composite materials. The R-CNN model was previously used to detect fibers in composite datasets including 2-dimensional slices from an X-Ray CT scan of a continuous SiC fiber reinforced SiC matrix mini-composite (single tow composite). With this approach we were able to place bounding boxes around the fibers with 98% accuracy. Recently we have employed the Mask R-CNN to both identify (i.e., apply boxes around) and segment (i.e., mask) the fibers. A python script using polygon mask annotations was written for rapid labeling to support the training processes. The machine learning models trained using the Mask R-CNN model achieved an average confidence score of 98.5%. |