Analysis of the RAFT Neural Network for Particle Image Velocimetry: Noise, Laser Sheet Effects, and Object Occlusion

Particle Image Velocimetry (PIV) experiments can efficiently generate large number of images in an experiment. Currently, processing these images using cross-correlation methods can be time consuming. As a result, convolutional neural networks (CNN) have gained popularity as an alternative method to process PIV image-pairs. In this work, we adopted a state-of-the-art CNN developed for optical flow applications and trained the CNN using computationally-generated PIV image-pairs and ground-truth flow fields. The computationally-generated PIV image-pairs used for training were designed to include flow features that are common in PIV experiments. We analyzed the robustness of the trained CNN to Gaussian noise added to the image, occlusion of the particles in the PIV image-pairs due to an object, and laser sheet lighting effects. We found that the CNN is robust to Gaussian noise in test image-pairs that contained noise-to-signal ratios 8 % above the noise-to-signal ratio included in training image-pairs. In addition, if the CNN was not trained on the object occlusion effects, the normalized error when tested on those effects was over 32.9 %. However, when the model was trained on this additional effect, test error dropped to approximately 3.9 %. With the laser sheet thickness effect, error dropped from 6.1 % to 4.2 % when trained on the effect. This result was close to error for models trained and tested on the dataset that were without the additional laser or object occlusion effects. So, CNN models could be trained to predict flows from PIV image-pairs that contained these additional effects.