Displacement extraction of background-oriented schlieren images using Swin Transformer

Abstract

Displacement extraction of background-oriented schlieren (BOS) is an essential step in BOS reconstruction, which directly determines the accuracy of the results. Typically, the displacement is calculated from the background images with and without inhomogeneous flow using the cross-correlation (CC) or optical flow (OF) method. This paper discusses the disadvantages of the CC and OF methods, and an end-to-end deep neural network was designed to estimate the BOS displacement. The proposed network is based on a Swin Transformer, which can build long-range correlations. A synthetic dataset used for training was generated using the simulated flow field by computational fluid dynamics. After training, the displacement can be obtained using the BOS image pair without additional parameters. Finally, the effectiveness of the proposed network was verified through experiments. The experiments illustrate that the proposed method performs stably on synthetic and real experimental images and outperforms conventional CC or OF methods and classic convolutional neural networks for OF tasks.