Method of Particle Field Reconstruction in Light Field Particle Image Velocimetry Based on Deep Residual Neural Networks
Mengxi Fu,Xiaoyu Zhu,Liang Zhang,Chuanlong Xu
DOI: https://doi.org/10.3788/aos240721
2024-01-01
Acta Optica Sinica
Abstract:Objective Light field particle image velocimetry (PIV) is a single- camera three-dimensional- dimensional flow field measurement method and has unique advantages under complex flow field measurement scenarios in narrow channels. The PIV technique consists of three parts: light field image acquisition, tracer particle spatial distribution reconstruction, and interrelated velocity field calculation. Among them, the reconstruction quality of the particle field will directly affect the accuracy and resolution of the velocity field measurement, which is the key link of light field PIV. Traditional particle field reconstruction methods for light field PIV, such as the joint algebraic reconstruction method, have low reconstruction efficiency, large computer memory requirement, and stretching effect of reconstructed particles. Optimization methods for traditional algorithms cannot completely solve the existing problems. Therefore, we introduce deep learning and propose a particle- reconstructed convolutional neural network (PRCNN) model based on deep residual neural networks to improve the quality and reconstruction efficiency of light field PIV particle field reconstruction. Methods Based on the geometric optics theory, we build an optical field imaging model and extract the optical field sub- aperture image containing information of multiple viewing angles from the original image of the optical field according to the optical field imaging characteristics. Meanwhile, the "three- dimensional spatial distribution of particles- optical field sub- aperture image" " dataset is constructed by numerical simulations. A deep residual neural network model is built, and a weighted MAE coupled reconstruction quality factor loss function is customized for training, with specific task objectives and data distribution characteristics taken into account. Further, the reconstruction quality and accuracy of the prediction model are evaluated by adopting numerical reconstruction methods, and the reconstruction efficiency is compared with that of the traditional SART algorithm. Finally, the measurement accuracy of the proposed method is analyzed in comparison with the traditional SART reconstruction algorithm by cylindrical bypass flow field measurement experiments. Results and Discussions The numerical reconstruction results show that in the tracer particle concentration range of 0.2-1.1, the reconstruction results of the proposed PRCNN model are all better than those of the SART algorithm, which improves the reconstruction quality factor by 153.83% (Fig. 11). Additionally, this model not only accurately reconstructs the true position of the particles, but also virtually eliminates the depth- direction stretching effect of the reconstructed particles (Fig. 12), improving the accuracy of particle position determination. Meanwhile, the proposed reconstruction method yields a reconstruction efficiency acceleration ratio of 3976.53 compared to the SART algorithm (Table 2), which can be employed for real-time- time particle field reconstruction. In the experimental evaluation results, by combining the 3D mutual correlation algorithm to calculate the 3D velocity field of the cylindrical winding flow field, the PRCNN model acquires the backflow behind the cylinder and demonstrates the staggered vortex structure (Fig. 15). At the central cross section of the pipe, the velocity results obtained by the PRCNN and SART algorithms are compared with the planar PIV measurements, and the velocity distribution results are basically the same (Fig. 16). After quantitatively comparing the velocity measurements at this plane, the average relative deviations of PRCNN and SART algorithms are 12.92 degrees o and 14.56 degrees o respectively, indicating that PRCNN can reconstruct a relatively accurate velocity field, with the computational error smaller than that of the SART algorithm. Finally, the practicality of the proposed method is verified. Conclusions To improve the optical field PIV 3D particle field reconstruction resolution and reconstruction efficiency, we propose a particle 3D distribution reconstruction method based on PRNCC, evaluate the reconstruction accuracy and efficiency of PRCNN by numerical simulations, and carry out an experimental evaluation study of cylindrical bypass flow field measurement. The results show that compared with the traditional SART algorithm, the reconstruction quality factor of the particle field of PRCNN is improved by 153.83 degrees o and the reconstruction time of a single light field image is only 0.025 degrees o of the SART algorithm. Additionally, the acceleration ratio reaches 3976.53, which proves that the proposed reconstruction method has high reconstruction quality and reconstruction efficiency. The reconstruction accuracy and efficiency of the cylindrical flow field measured by PRCNN and the traditional SART algorithm are evaluated by numerical simulations. The three-dimensional- dimensional flow field of cylindrical flow measured by PRCNN and traditional SART algorithms is consistent, and the velocity distribution in the central cross section of the pipe (Z/D=0) is basically the same as that measured by planar PIV. Meanwhile, the average relative errors of PRCNN and SART measurements are 12.92 degrees o and 14.56 degrees o respectively, which indicates that the PRCNN method can be utilized for accurate three-dimensional- dimensional flow field measurements by the light- field PIV technique, thus realizing accurate 3D flow field measurements.