Deep learning approach for flow visualization in background-oriented schlieren

Diffractive optical element-based background-oriented schlieren (BOS) is a popular technique for quantitative flow visualization. This technique relies on encoding spatial density variations of the test medium in the form of an optical fringe pattern; and hence, its accuracy is directly influenced by the quality of fringe pattern demodulation. We introduce a robust deep learning-assisted subspace method, which enables reliable fringe pattern demodulation even in the presence of severe noise and uneven fringe distortions in recorded BOS fringe patterns. The method's effectiveness in handling fringe pattern artifacts is demonstrated via rigorous numerical simulations. Furthermore, the method's practical applicability is experimentally validated using real-world BOS images obtained from a liquid diffusion process.