Balanced deep learning-based bubble segmentation: Model comparison, optimization, and application in microbubble detection

Abstract

The accurate segmentation and analysis of bubbles are crucial for understanding bubble generation mechanisms and improving industrial microbubble detection. This study aims to evaluate and optimize deep learning-based bubble segmentation models. Firstly, a systematic model evaluation matrix is proposed, including the general model performance, defocused bubble size prediction accuracy, and overlapping bubble segmentation. Secondly, four models, including SplineDist, StarDist, YOLOv8-seg, and Mask R-CNN, are compared. The SplineDist-M16 model demonstrates superior image processing speed (7.84 FPS) and high accuracy in bubble size prediction with minimal misdetection (6.1 %). Compared to other models, SplineDist-M16 excels in edge fitting and overlapping bubble identification. The optimized model provides rapid, accurate measurement of bubble quantity, size, and shape, offering insights into bubble formation and guiding microbubble generator design. This study paves the way for real-time microbubble detection in industrial applications and suggests further model improvements through simulated data training and enhanced overlapping bubble segmentation. Furthermore, the SplineDist-M16 model was utilized to analyse the impact of flow rate and backpressure on microbubble characteristics generated by a Venturi-tube microbubble generator. The results show that increased flow rate reduces bubble size and increases bubble circularity, while backpressure has minimal impact on bubble size distribution and shape.