Fast and accurate approach 3D-micro PTV localization in densely seeded flows by deep convolutional neural network

In 3D-microPTV the depth of the particle is inferred through changes in its image. Algorithms developed to determine the depth of a particle from its image are both computationally expensive and limited in the density of particles in an image which they can analyze. This, ultimately limits the temporal resolution with which 3D fluid flow can be visualized. Here we developed an end-to-end convolutional neural network to analyze densely seeded microflows. Convolutional Neural Network (CNN) has proven to be a powerful image-processing tool in computer vision applications such as pattern recognition, object localization, and object tracking. In this study, we took the advantage of CNN and presented an end-to-end model to localize sub-10nm objects in 3 dimensional. The performance of the model compared against the MLE algorithm as the gold standard using the experimental results of BBM system. As a result of the comparison, the lateral and axial localization precisions using CNN are convincingly close to the MLE algorithm results and insensitive to the background noise. Meanwhile, the computation runtime of the CNN localization model is greatly reduced and the number of particles capable of being simultaneously localized is increased.