| About this Abstract |
| Meeting |
MS&T23: Materials Science & Technology
|
| Symposium
|
Additive Manufacturing: Design, Materials, Manufacturing, Challenges and Applications
|
| Presentation Title |
Artificial Grain Image Generation for CNN-Based Segmentation Training in Additively Manufactured Components |
| Author(s) |
Peter Warren, Md Shahjahan Hossain, Pranta Sarkar, Asher Perez, Daniel Homa, Gary R. Pickrell, Ranajay Ghosh, Ramesh Subramanian, Jayanta Kapat, Navin Manjooran |
| On-Site Speaker (Planned) |
Navin Manjooran |
| Abstract Scope |
This paper presents a novel approach to training a Convolutional Neural Network (CNN) for grain analysis of 3D printed components without the need for any real data. The analysis of grain size is crucial in certifying the quality of metals and ceramics, ensuring that the components meet the required standards. Traditional methods often rely on manual measurements, which are time-consuming and prone to human error. In this study, a grain data generation technique is proposed that eliminates the dependency on real data, enabling efficient and accurate grain segmentation. The proposed method involves measuring one or two representative grains and constructing a Voronoi diagram plot. To enhance the realism of the artificial dataset, various imperfections such as surface scratches, pores, and color variations are introduced. Furthermore, the dimensions of the Voronoi plot are adjusted to closely resemble the grains of the specific material being analyzed. This allows a comprehensive and customizable dataset to be generated and tailored towards the desired grain geometries that are under consideration. |