Oil-flow Convolutional Neural Network

This site supports the paper published [paper], showcasing a Convolutional Neural Network (CNN) trained to predict the flow direction from an oil flow visualizations.

Overview

The code provided in this repository shows an example to predict the flow direction.

Example visualization:

The image shows an oil-flow visualization over a backward facing ramp in gray scales.

Output prediction:

The following image shows the predicted direction (arrows) obtained by the CNN. The color of each arrows depends on the outlier algorithm output; red is not calculated as an outlier, blue is an outlier.

Step 1 - corrected values:

The first step to correct the field is made by rotating each outlier by 180° and calculating its neighbourhood to check if the change has corrected the outlier.

Step 2 - corrected values:

The second step takes into account the neighbourhood of each outlier to correct the direction. The new direction is an average of its neighbours.

References

This project uses an implementation of the algorithm described in:

Jerry Westerweel and Fulvio Scarano, "Universal outlier detection for PIV data", Experiments in Fluids, 2005 DOI: https://doi.org/10.1007/s00348-005-0016-6