| Abstract Scope |
With the growing applications of cold-spray technology, the ability to quickly analyze and compare powder qualities has become increasingly important. Previous materials science and engineering research has demonstrated Mask R-CNN (Convolutional Neural Network) models' capability to meet this need, distinguishing satellites from their corresponding parent-particle within an image. However, this was performed on a homogeneous dataset with only slight variation, consisting of only one powder type collected at a single magnification. The present work builds upon this, providing the following contributions: 1) demonstrates a Mask R-CNN can be implemented on a diverse dataset of multiple powders at various magnifications, 2) provides a tool to convert model predictions into image annotations, allowing a partially-automated process to grow a dataset rapidly, 3) highlights potential reproducibility issues within Mask R-CNN models due to inherent nondeterminism and providing possible resolutions. |