About this Abstract |
Meeting |
TMS Specialty Congress 2024
|
Symposium
|
2nd World Congress on Artificial Intelligence in Materials & Manufacturing (AIM 2024)
|
Presentation Title |
Grain Segmentation Refinement Based on Perceptual Grouping Principles Using Conditional Random Fields |
Author(s) |
Doruk Aksoy, Tim Rupert, William Bowman |
On-Site Speaker (Planned) |
Doruk Aksoy |
Abstract Scope |
Accurate segmentation of grains in electron microscopy images is key for high-throughput microstructure analyses that determine material properties. However, the task is challenging due to factors like intricate network connectivity influenced by crystallography and local atomic environments. Traditional post-processing of machine vision-generated segmentation masks is complex, especially when dealing with interconnected linear structures. Techniques like convolutional neural networks often miss fine details due to their large receptive fields. To address these issues, we introduce a generalizable post-processing approach that enhances segmentation masks in grain boundary networks. This approach uses conditional random fields and leverages perceptual grouping principles such as similarity, proximity, closure, and continuity. We validated the method's effectiveness using segmentation masks generated by a U-Net architecture on transmission electron microscopy images of an electrically conducting polycrystalline oxide. The approach demonstrated significant improvements in accuracy across multiple evaluation metrics, suggesting its utility for segmentation mask refinement in various application scenarios. |
Proceedings Inclusion? |
Definite: Other |