Instance Segmentation And Flow-Regime-Based Classification Of Tufts Using A Deep Neural Network

Abstract

Tufts are often used to identify unsteady flow and flow separation regions on aerodynamic models. To deduce information about the flow, the tufts' range of motion can be inspected visually in a sequence of short-exposure frames or on long-exposure images . We propose using Mask R-CNN, a Convolutional Neural Network (CNN) for instance segmentation, to segment and classify tufts as stationary or fluttering based on their range of motion. This method automates the detection of regions of unsteady flow while at the same time providing accurate segmentation masks and endpoint predictions for each tuft. The network is trained on fully synthetic data, and we obtain good results in the segmentation, classification, and keypoint detection of both synthetic and experimental data. The proposed method demonstrates the value of employing neural networks for vision-based tasks in experimental fluid dynamics and highlights the potential of using synthetic training data for similar applications.