Two-Phase Flow Image Analysis Using Pre-trained Segmentation Anything Model
POSTER
Abstract
Measuring sizes and shapes of bubbles, droplets, and cells in transport quickly and consistently is important in Biomedical, Pharmaceutical, and Tribological applications. Classic computer‑vision (CV) performance in this context is impacted greatly by changes in image quality, lighting conditions, etc.
SAM2 is a strong pre-trained model that can segment many kinds of objects, where performance depends on prompt quality (points/boxes) and the ability to filter the masks it returns. Using bubbles visualized in the context of engine lubricant flow simulated on a bench top facility, we demonstrate a simple, reusable framework for tracking bubbles over time using lightweight, classic CV to auto‑create prompts from each frame; running SAM2 to get candidate masks and clean the results with SAM2’s own mask‑quality scores.
While specific parameters for prompt generation and mask filtering may vary from application to application, we demonstrate the effect of pre‑/post‑processing choices against a manually generated ground truth dataset using standard metrics—precision, recall, F1, Dice, and IoU. On aerated‑lubricant flow visualization videos, this workflow produces accurate and repeatable bubble sizes with little manual intervention and significantly reduced processing time compared to base model prompting. Because it relies on proven CV steps and model‑provided scores, the same approach can be transferred to the aforementioned applications.
SAM2 is a strong pre-trained model that can segment many kinds of objects, where performance depends on prompt quality (points/boxes) and the ability to filter the masks it returns. Using bubbles visualized in the context of engine lubricant flow simulated on a bench top facility, we demonstrate a simple, reusable framework for tracking bubbles over time using lightweight, classic CV to auto‑create prompts from each frame; running SAM2 to get candidate masks and clean the results with SAM2’s own mask‑quality scores.
While specific parameters for prompt generation and mask filtering may vary from application to application, we demonstrate the effect of pre‑/post‑processing choices against a manually generated ground truth dataset using standard metrics—precision, recall, F1, Dice, and IoU. On aerated‑lubricant flow visualization videos, this workflow produces accurate and repeatable bubble sizes with little manual intervention and significantly reduced processing time compared to base model prompting. Because it relies on proven CV steps and model‑provided scores, the same approach can be transferred to the aforementioned applications.
*This work was carried out at Vanderbilt Aerospace Design Laboratory through a summer internship at the Vanderbilt School of Engineering.
Presenters
-
Zeyu Zhao
- Mount Holyoke College