A lightweight convolutional neural network to reconstruct deformation in BOS recordings

Abstract

We introduce a Convolutional Neural Network (CNN) that is specificallydesigned and trained to post-process recordings obtained by BackgroundOriented Schlieren (BOS), a popular technique to visualize compressibleand convective flows. To reconstruct BOS image deformation, we deviseda lightweight network (LIMA) that has comparatively fewer parameters totrain than the CNNs that have been previously proposed for optical flow.To train LIMA, we introduce a novel strategy based on the generation of synthetic images from random-irrotational deformation fields, which areintended to mimic those provided by real BOS recordings. This allows us to generate a large number of training examples at minimal compu-tational cost. To assess the accuracy of the reconstructed displacements,we consider test cases consisting of synthetic images with sinusoidal dis-placement as well as images obtained in the experimental studies of a hotplume in air and a flow past and inside a heated hollow hemisphere. By comparing the reconstructed deformation fields using the LIMA or conven-tional post-processing techniques used in Direct Image Correlation (dic)or conventional image cross-correlation, we show that LIMA is more accu-rate and robust in the synthetic test case. When applied to experimentalBOS recordings, all methods provide similar and consistent deformation fields. As LIMA is capable of achieving a comparable or better accuracyat a fraction of the computational costs, it represents a valuable alternative to conventional post-processing techniques for BOS experiments.

A quantitative reconstruction method for background-oriented schlieren by reducing the weight matrix

“…However, the basic assumptions of optical flow may affect the accuracy of the reconstruction, as optical flow is sensitive to changes in brightness and tends to capture small displacements. Moreover, this method cannot employ the advantages of using techniques like deep learning for displacement extraction, because the deep learning has been proven to achieve higher accuracy in BOS displacement extraction 12,13 .…”

A lightweight neural network designed for fluid velocimetry

“…To assess the reconstruction error we compare with the ground truth for the synthetic test cases, whereas for the experimental test cases we employ an a-posteriori uncertainty quantification (UQ) by image matching [19]. This work is part of a wider effort to devise ML diagnostic tools for experimental fluid dynamics, which includes also the application of LIMA in Background Oriented Schlieren (BOS) [20] .…”