Materials Data Science and Informatics (IAS-9) · Forschungszentrum Jülich
Automated segmentation and tracking of grain growth in TEM data
The in-situ TEM videos contain hundreds to thousands of frames, making manual annotation and identity tracking slow and impractical. I am developing a pipeline that moves from single-frame segmentation to persistent grain identities so researchers can quantify area, morphology, boundary motion, and grain growth over time.
Step 01 · Single frame
Grain segmentation
Convert difficult and changing microscopy contrast into grain-level masks that establish the objects to be tracked.
- In-situ TEM
- Instance segmentation
- SAM · SAM2 · SAM3
- Prompt engineering
- MatSAM
- U-Net
- EBSD
Step 02 · Time series
Video object tracking
Preserve each grain’s identity through boundary motion, growth, shrinkage, and changing neighbours, so its evolution can be quantified.
- Video object tracking
- Persistent IDs
- Hungarian algorithm
- SAM2 · SAM3 video segmentation
- Prompt engineering
- PyTorch · OpenCV
- YOLO — planned
