BubSAM: Bubble segmentation and shape reconstruction based on Segment
Anything Model of bubbly flow
Abstract
Accurate detection and analysis of bubble size and shape in bubbly flow
are critical to understanding mass and heat transfer processes.
Convolutional neural networks have limitations in different bubble
images due to their dependence on large amounts of labeled data. A new
foundational Segment Anything Model (SAM) recently attracts lots of
attention for its zero-shot segmentation performance. Herein, we
developed a novel image processing method named bubSAM, which achieves
efficient and accurate bubble segmentation and shape reconstruction
based on SAM. The segmentation performance of bubSAM is 30% higher than
that of SAM, and its accuracy reaches 90% under different bubbly flow
conditions. The accuracy of bubble shape reconstruction algorithm (BSR)
in bubSAM is about 30% higher than that of typical ellipse fitting
method, thus better restoring the geometric shape of bubbles. BubSAM can
provide great support for understanding gas-liquid multiphase flow and
design of industrial multiphase reactors.