Bubble velocimetry using the conventional and CNN-based optical flow algorithms

In the present study, we introduce new bubble velocimetry methods based on the optical flow, which were validated (compared) with the conventional particle tracking velocimetry (PTV) for various gas–liquid two-phase flows. For the optical flow algorithms, the convolutional neural network (CNN)-based models as well as the original schemes like the Lucas-Kanade and Farnebäck methods are considered. In particular, the CNN-based method was re-trained (fine-tuned) using the synthetic bubble images produced by varying the density, diameter, and velocity distribution. While all models accurately measured the unsteady velocities of a single bubble rising with a lateral oscillation, the pre-trained CNN-based method showed the discrepancy in the averaged velocities in both directions for the dilute bubble plume. In terms of the fluctuating velocity components, the fine-tuned CNN-based model produced the closest results to that from PTV, while the conventional optical flow methods under- or over-estimated them owing to the intensity assumption. When the void fraction increases much higher (e.g., over 10%) in the bubble plume, the PTV failed to evaluate the bubble velocities because of the overlapped bubble images and significant bubble deformation, which is clearly overcome by the optical flow bubble velocimetry. This is quite encouraging in experimentally investigating the gas–liquid two-phase flows of a high void fraction. Furthermore, the fine-tuned CNN-based model captures the individual motion of overlapped bubbles most faithfully while saving the computing time, compared to the Farnebäck method.