A flexible image processing technique for measuring bubble parameters based on a neural network

Abstract

The main parameters in a liquid-gas system are the size, shape, and kinetic parameters of bubbles, which affect the mass transfer process. To describe the mass transfer process accurately, it is necessary to characterize bubble parameters such as bubble volume and surface area under different experimental conditions. However, these bubble parameters are not accurately assessable by conventional methods, which are limited by poor robustness and flexibility, including numerical methods and conventional neural networks. Based on the neural network technique, we propose a flexible method for measuring bubble parameters in three steps: segmentation of bubble images, segmentation of overlapping bubbles, and extraction of bubble feature parameters. This technique combines geometric, optical, and topological information. To verify the effectiveness of our method, we considered a gas-liquid reactor and microbubble radiosensitizer for brachytherapy as testing scenarios and used several groups of experimental images and synthetic images as testing sets. In the first step, the accuracy of the segmentation of bubble images was greater than 90% in all scenes and reached 99% in some scenes. In the second step, the proposed method was capable of segmenting overlapping bubbles and avoiding excessive segmentation. In the final step, our approach considers the geometric shape and deflection angle of bubbles. Therefore, the bubble parameters of largely deformed and deflected bubbles were obtained with high flexibility. Our method is a powerful tool with robustness and flexibility for the study of gas-liquid interface measurements in chemical and biochemical processes.