A deep learning-based image processing method for bubble detection, segmentation, and shape reconstruction in high gas holdup sub-millimeter bubbly flows

The sub-millimeter bubble technique can enhance the gas–liquid inter-phase mass transfer by significantly reducing the bubble size and increasing the gas–liquid interfacial area. To accurately describe the flow and mass transfer characteristics, it is necessary to characterize bubble parameters. High-speed photography followed by image processing is an effective way to characterize the gas bubbles in the multiphase flows. However, the efficient image processing method for the sub-millimeter bubbly flows with high gas holdup and high bubble overlap has not been reported yet. The present work developed a novel deep learning-based image processing method for bubble detection, segmentation, and shape reconstruction in high gas holdup sub-millimeter bubbly flows. In order to segment the highly overlapping sub-millimeter bubbles, our method was built based on Mask R-CNN, with which the pixel-level segmentation masks can be obtained, and the shape of the overlapping bubbles can be accurately described. The feature pyramid architecture was coupled with ResNet101 and Feature Pyramid Network to detect sub-millimeter bubbles with significant size differences. A shape reconstruction module was proposed to restore the real shape of overlapping bubbles and improve prediction accuracy. In order to sufficiently validate the proposed method, adequate images of sub-millimeter bubbly flows were obtained by changing the experimental media (air-tap water, air-sodium dodecyl sulphate aqueous solution, air-diesel, and air-diesel-fine catalyst particles), reactor configurations (3D beds and 2D beds), lenses, and photography (shadowgraphy and front illumination). Our method shows high accuracy under the experimental conditions and can process sub-millimeter bubble images under gas holdup up to 20%.