Coaxing Truth from Noise: A Pipeline for Implementing Spatially Aware Filtering Algorithms on Velocity Fields
Mentor 1
Roshan D'Souza
Location
Union 250
Start Date
5-4-2019 12:40 PM
Description
Particle Image Velocimetry (PIV) is an experimental method that uses high-speed imaging and laser-illuminated particles to generate a vector field which can be used to verify mathematical or computational fluid models. However, PIV is very sensitive to optical occlusions and particle density, and in these cases will generate noisy data that can be near useless for verification purposes; these problems become exponentially worse in volumetric (3D) PIV. In order to separate the noise from the truth, several filtering algorithms were developed. The first algorithm processes the images and removes the static image background using Singular Value Decomposition (SVD) in a hands-free approach, the rest of the filters are implemented in 3D and run in sequence; these algorithms can be modified with multiple parameters for best results. The spatial filters are built on the K-Nearest Neighbors algorithm and statistical techniques, each one specializes at removing certain types of spurious vectors. These algorithms were tested on data from brain aneurysm models, where there were many particles stuck to the model walls and stent structures. The initial SVD filter was able to subtract the bright stuck particles, as well as the reflections and discontinuities to reveal the particles in high contrast. The pipeline filters were then able to take the reconstructed randomly sampled vector data and filter out erroneous vectors, this clean data was then resampled into a structured grid that can be displayed in post-processing software. This data was significantly cleaner, and the flow streamlines were co-linear implying the filters had removed all the data that caused divergence in the flow. This filtering pipeline shows promising results, and could significantly improve the state of the art in PIV experiments where reflections, optical occlusions, or stuck particles become major issues, and also in volumetric PIV where particle density causes serious noise error.
Coaxing Truth from Noise: A Pipeline for Implementing Spatially Aware Filtering Algorithms on Velocity Fields
Union 250
Particle Image Velocimetry (PIV) is an experimental method that uses high-speed imaging and laser-illuminated particles to generate a vector field which can be used to verify mathematical or computational fluid models. However, PIV is very sensitive to optical occlusions and particle density, and in these cases will generate noisy data that can be near useless for verification purposes; these problems become exponentially worse in volumetric (3D) PIV. In order to separate the noise from the truth, several filtering algorithms were developed. The first algorithm processes the images and removes the static image background using Singular Value Decomposition (SVD) in a hands-free approach, the rest of the filters are implemented in 3D and run in sequence; these algorithms can be modified with multiple parameters for best results. The spatial filters are built on the K-Nearest Neighbors algorithm and statistical techniques, each one specializes at removing certain types of spurious vectors. These algorithms were tested on data from brain aneurysm models, where there were many particles stuck to the model walls and stent structures. The initial SVD filter was able to subtract the bright stuck particles, as well as the reflections and discontinuities to reveal the particles in high contrast. The pipeline filters were then able to take the reconstructed randomly sampled vector data and filter out erroneous vectors, this clean data was then resampled into a structured grid that can be displayed in post-processing software. This data was significantly cleaner, and the flow streamlines were co-linear implying the filters had removed all the data that caused divergence in the flow. This filtering pipeline shows promising results, and could significantly improve the state of the art in PIV experiments where reflections, optical occlusions, or stuck particles become major issues, and also in volumetric PIV where particle density causes serious noise error.
