Deep-Learning Based Three-Dimensional Analysis of Bubble Flows From Light Field Images

Abstract

An accurate measurement for the modality and motion parameters of bubbles is of great significance. In the existing bubble measurement methods, there are some problems that are desirable to be solved, such as system complexity, poor real-time performance, and the deviation due to two-dimensional measurement. To overcome these challenges, we developed a novel three-dimensional analysis method based on light field imaging diagnosis and deep learning algorithm. Different from traditional two-dimensional reconstruction, the bubble depth can be computed from light field images directly through digital refocusing technology. After calibrating, the conversion between the camera coordinate system and the real-world coordinate system is achieved through the sharpness evaluation algorithm. According to the corresponding relationship, the refocused image could be calibrated to the actual position in real world. Combined with the Multi-input Residual Convolution Neural Network, (MRCNN), the bubble depth could be computed fully automated from given images at high accuracy. Based on the above works, the three-dimensional reconstruction model for bubble flow can be established by coupling the depth and the parameters extracted through bounding boxes. The proposed method solves the problem that conventional imaging can only perform the two-dimensional measurement, which contributes to the error during the measurement process. Results show a promising performance on the three-dimensional reconstruction of bubble flow, validating the feasibility of the three-dimensional measurement method for bubbles in gas-liquid two-phase flow based on light field imaging diagnosis and deep learning algorithm.