文献类型：期刊文献

英文题名：Accurate bubble identification and segmentation in gas-liquid two-phase flows using pretrained neural network models in cellpose

作者：Yu, Bo[1];Wu, Tao[1];Chen, Feiyu[1];Yang, Zixian[1];Yuan, Fang[1];Yang, Qiang[1];Liu, Bo[1]

机构：[1]East China Univ Sci & Technol, Dept Mech & Power Engn, Shanghai 200237, Peoples R China

年份：2026

卷号：110

外文期刊名：FLOW MEASUREMENT AND INSTRUMENTATION

收录：;EI(收录号：20261420433876);WOS:【SCI-EXPANDED(收录号:WOS:001740338400001)】；

基金：This work was financially supported by the National Key Research and Development Program of China (Grant.No.2022YFE0130000-03) , the National Natural Science Foundation of China (Grant. Nos.52025103,22178099) , the Shanghai Natural Science Foundation (Grant.No.21ZR1417000) , the Shanghai Pilot Program for Basic Research (Grant. No. 22TQ1400100-11) .

语种：英文

外文关键词：Bubble recognition; Cellpose algorithm; Gas-liquid two-phase flow; Image segmentation

摘要：Accurate determination of bubble size, morphology, and spatial distribution is essential for quantitative analysis of gas-liquid two-phase flows, such as bubble columns and water electrolyzers. While deep learning has shown strong potential for bubble image analysis, most methods require large annotated datasets. In this paper, we adapt pretrained neural network models from Cellpose, originally developed for cell segmentation, for efficient bubble recognition using limited training data. Quantitative analysis shows that for millimeter-scale single bubbles, the Cyto model achieves a Dice coefficient over 0.9 and an average relative error of about 3% in equivalent diameter, enabling accurate trajectory tracking. For micrometer-scale overlapping bubbles, the Cpsam model trained on just 60 samples reaches over 90% in Dice coefficient, precision and recall, with bubble count deviation within +/- 5 and diameter error below 5%. Trained on only 15 images, it also segments bubbles well in complex scenarios like bubble columns and alkaline water electrolysis for hydrogen production. These results demonstrate that the proposed Cellpose-based approach enables accurate and efficient bubble image analysis without extensive labeled datasets or complex hardware, providing a robust tool for quantitative characterization of gas-liquid two-phase flow systems.