Automatic extraction of wall streamlines from oil-flow visualizations using a convolutional neural network

Abstract Oil-flow visualizations represent a simple means to reveal wall streamline patterns. Yet, the evaluation of such images can be a time-consuming process and is subjective to human perception. In this article, we present a fast and robust method to obtain quantitative insight based on qualitative oil-flow visualizations. Specifically, the local wall streamline direction is predicted by a convolutional neural network. The supervised training of this network was based on an extensive dataset involving approximately one million image patches that cover variations of the flow direction, the wall shear-stress magnitude and the oil-flow mixture. For a test dataset that is distinct from the training data, the mean prediction error of the flow direction is as low as three degrees. A reliable performance is also noted when the model is applied to oil-flow visualizations obtained from the literature, demonstrating the generalizability required for an application in diverse flow configurations. The trained model is available at https://github.com/AeroTUBerlin/OilFlowCNN .