Convolutional-SMART: a fast reconstruction technique for tomographic PIV

Abstract

Tomographic reconstruction, a critical process for tomographic particle image velocimetry (Tomo-PIV), remains inefficient due to the required massive memories and high computational cost. In this work, a fast tomographic reconstruction technique is proposed to improve the efficiency significantly. The weighting coefficient, which represents the contribution of the voxel to the corresponding pixel intensity, is remodeled to be independent of the voxel’s positions by artificially improving the particle image resolution. Consequently, the simultaneous multiplicative algebraic reconstruction technique (SMART) is re-implemented with convolution operations. The proposed method is named Convolutional-SMART (Conv-SMART). Moreover, the numerous convolution operations are accelerated using a graphics processing unit (GPU) to further reduce the reconstruction time. A synthetic three-dimensional 3D experiment with a vortex ring is carried out to numerically evaluate the precision and efficiency of the proposed method. The results show that the speed-up ratio of Conv-SMART to the original SMART reaches about five times faster in the central processing unit (CPU) environment and 15 times faster in the GPU environment without losing accuracy when the particle density is 0.05 particles per pixel (ppp) and the resolution is 20 voxels/mm. The speed-up ratio as a function of the particle density and resolution is also provided. Conv-SMART is also applied to the left ventricular Tomo-PIV experiment. The velocity field derived from Conv-SMART is consistent with that from SMART, whereas Conv-SMART achieves 50 times faster within the GPU.