Volumetric reconstruction of flow particles through light field particle image velocimetry and deep neural network

Abstract

Tomographic reconstruction of three-dimensional (3D) tracer particle distributions through light field particle image velocimetry (LF-PIV) faces challenges in low reconstruction resolution owing to the elongation effect and extensive computational cost incurred by the iterative process. To resolve these challenges, this study proposes a deep neural network-based volumetric reconstruction approach to alleviate the reconstruction elongation and enhance the reconstruction efficiency. A tailored deep learning model (namely, LF-DNN) incorporating residual neural network architecture and a novel hybrid loss function is established to reconstruct the particle distributions through LF images. The parallax information of the flow field decoded from the raw LF data is leveraged as the input features of the network model. Comparative studies between the proposed method and the traditional tomographic reconstruction algorithms (multiplicative algebraic reconstruction technique, MART and pre-recognition MART, PR-MART) are performed through synthetic datasets. Experiments on a cylinder wake flow are further conducted to validate the performance of the proposed LF-DNN. The results indicate that the LF-DNN outperforms MART and PR-MART in terms of the reconstruction quality, mitigation of elongation effect, and noise resilience. The LF-DNN also improves the reconstruction efficiency which is 9.6 and 7.1 times higher than the MART and PR-MART, respectively. The relative error of the cylinder wake flow achieved by the LF-DNN is 2% lower than the MART. It suggests that the LF-DNN can facilitate accurate volumetric particle reconstruction and hence the three-dimensional flow measurement by single camera-based LF-PIV.