Deep-learning-based image preprocessing for particle image velocimetry

Abstract

Particle image velocimetry (PIV) is an important measurement technique for estimating global fluid motion by capturing particle motion from a pair of images. Hence, obtaining high-quality particle images is critical for a PIV estimation. The purpose of this study is to propose a deep-learning-based technique for PIV image preprocessing. Specifically, we first designed a deep convolutional network called Bilateral-CNN for a pair image preprocessing task, which embeds residual learning and batch normalization. Bilateral-CNN fully considers the feature connections between image pairs. Second, to train our model, we generate a synthetic dataset that includes Gaussian noise and various light intensities. Considering a real fluid scene, different background information such as rising bubbles is superposed into the images, which makes the particle images more challenging. Finally, we conducted experiments on synthetic and experimental particle images. Extensive results indicate that our method can effectively overcome the particle image interference noise under different situations. In addition, the processed particle images are calculated using cross-correlation WIDIM algorithm, and a high-precision velocity field is obtained by fast inference speed.