En
期刊文章
文章精选
当期目录
过刊浏览
高被引
高下载
特邀文章
专刊合集
待刊文章
Volume 113
您当前的位置：首页 > 期刊文章 > 过刊浏览 > Volumes 108-119 (2025) > Volume 113
Comparative study of bubble segmentation methods for image analysis in hydrogen-based fluidized beds
Chuanhao Wang, Yuchen Liu, Jiehan Zhang, Shiyuan Li *
School of Energy and Environmental Engineering, University of Science and Technology Beijing, Beijing, 100083, China
10.1016/j.partic.2026.03.014
Volume 113, June 2026, Pages 88-101
Received 13 July 2025, Revised 27 February 2026, Accepted 4 March 2026, Available online 25 March 2026, Version of Record 1 April 2026.
E-mail: lishiyuan@ustb.edu.cn
Highlights

• DeepLabV3+ framework is developed for simultaneous gas–solid phase segmentation in iron ore powder fluidization.

• Segmentation accuracy and efficiency of DeepLabV3+ with different backbone networks are systematically evaluated.

• The method is compared with threshold-based approaches and the SAM3 foundation model.

• Fluidization behaviors of different particles under various conditions are analyzed using the proposed method.


Abstract

Clarifying the fluidization behavior of iron ore powders in fluidized beds is a fundamental issue for realizing hydrogen-based fluidized bed reduction. Digital image analysis (DIA), as a promising non-intrusive diagnostic technique, provides an effective approach for the quantitative characterization of fluidization behavior. In this study, a DeepLabV3+-based deep learning framework was employed to simultaneously segment the gas and solid phases during the fluidization process, and the effects of different backbone network configurations on training cost and segmentation performance were systematically evaluated. On this basis, the proposed method was further compared with conventional threshold-based segmentation methods and the general-purpose foundation model SAM3. The results indicate that using ResNet18 as the backbone achieves a favorable balance between segmentation accuracy and computational efficiency. The deep learning–based segmentation method shows strong adaptability to complex backgrounds and diverse fluidization states, and its segmentation accuracy and processing efficiency are significantly superior to those of traditional image segmentation methods and SAM3. Further experimental results based on this approach reveal that, at the same fluidization number, the iron ore powder system generates smaller bubbles with lower rising velocities, exhibiting a bubble population characterized by smaller sizes, higher number density, and lower rising velocities. With increasing hydrogen concentration in the fluidizing gas, the bubble size decreases further and the bed expansion ratio is reduced, while the variations in bubble shape and rising velocity remain insignificant.

Graphical abstract
Keywords
Gas-solid two-phase flow; Hydrogen metallurgy; Quasi-2D fluidized bed; Deep learning; Image segmentation; Digital image
文章导读