Complex-background-oriented schlieren technology based on convolutional neural networks

Background-oriented schlieren (BOS) is widely used for visualizing transparent flow fields, but traditional displacement extraction algorithms struggle in large-scale flow fields with complex, noisy backgrounds. This paper presents robust-IRR model, based on IRR-PWCNet, which improves robustness through physics-based noise augmentation. Experimental results demonstrate that robust-IRR reduces endpoint error (EPE) by 43.06% in noisy remote sensing backgrounds compared to other methods, while maintaining high computational efficiency. Moreover, it reduces the error by 11.72% to 14.20% at different PSNR levels compared to the IRR without noise enhancement. Furthermore, to reconstruct the flow field and wavefront shapes, we apply a bi-conjugate gradient (BiCG) method with ASDI-based initial values, achieving a 7.47% error reduction over using BiCG or ASDI alone. The proposed method offers a robust and efficient solution for high-precision flow field measurement, visualization, and wavefront reconstruction in complex environments.