DualSight: multi-stage instance segmentation framework for improved precision

Abstract Powder properties, particularly their morphology (which includes size, shape, surface roughness, etc.), play a critical role in the quality of cold spray additively manufactured materials. A change in feedstock powder morphology can impact flowability and deposition quality, deposition efficiency, and porosity. Image analysis can be used to quantify a powder’s morphology, but performing this analysis with manual visual inspection can be laborious and time-consuming. Alternatively, computer vision techniques have shown promise in automating powder morphology analysis, reducing the work and time required to quantify a powder’s morphology. However, the capabilities of these models are limited to the quality of their training data, which can be equally difficult and expensive to collect and annotate. Thus, this work presents DualSight, a novel multi-stage computer vision framework that improves metallic powder segmentation quality without requiring additional data or model training. With this framework, powder morphology can be extracted more accurately from scanning electron microscope images, enabling a more informed manufacturing process.