Author(s): Chenxi Liang; Philippe Gourbesville; Qiang Ma
Linked Author(s): Philippe Gourbesville, Qiang Ma
Keywords: Image velocimetry Deep learning Large-scale particle image velocimetry (LSPIV) Recurrent All-Pairs Field Transforms (RAFT)
Abstract: Image velocimetry has shown broad application prospects in many fields such as flow velocity monitoring and fluid dynamics research. Based on a series of physical experiments under different scenarios with various flow velocities, tracer states, illumination conditions, and resolutions, a high quality database has been constructed and applied for a deep learning algorithm named with Fusion-UNet for training. The Fusion-UNet algorithm is relied on the architectural characteristics of U-Net, with this kind of architecture, the output of the algorithm is able to reach the operational request of having reasonable flow velocity in real time. This algorithm fuses adjacent frames, combines the powerful feature extraction capabilities of convolutional neural networks, accurately captures subtle changes between adjacent frames, and achieves efficient processing of time dimension data in an indirect way. The testing results show that this algorithm shows extremely stable performance under all different scenarios with only 5% error compared to the measured velocities which is obviously more accurate than other classical approaches such as LSPIV and RAFT in image velocimetry field. The innovative approach presented in this study shows high applicability to be promote in the scenarios of flash flood or river flood measurement to obtain the real-time data within certain accuracy.
Year: 2025