Random Walks for Vector Field Denoising João Paixão, Marcos Lage, Fabiano Petronetto, Alex Laier, Sinésio Pesco, Geovan Tavares, Thomas Lewiner, Hélio Lopes Matmidia Laboratory – Department of Mathematics PUC–Rio – Rio de Janeiro, Brazil
Motivation Vector Fields in Science and Engineering Flow in an artificial heart Flow patterns in a tube University of Cambridge (2009)
Motivation Noise in vector data-acquisition Flow around a live swimming fish (Yoshida et al 2004)
Problem
Problem:Noise Denoising
Gaussian Filtering E.g. 5x5 Gaussian Filter
Limitations Feature Destruction
Limitations Feature Destruction
Random Walks on the Graph Feature
Previous Work Smolka et al Random Walk for Image Enhancement
Previous Work Sun et al Mesh Denoising
Random Walks for Vector Fields What we want -Meshless -Feature-preserving What do we need -Graph -Probabilities that avoid crossing features
How to build the graph
Feature Functions Direction Magnitude
Feature Functions Direction Magnitude Other feature functions in the paper!
Probabilities is the neighborhood of vector i Probability from vector i to vector j
Time to walk A B
A B
A B
A B
A B
- the probability of going from node A to node B after n steps A B
Random Walk Filtering Weighted Average of Random Walk Probabilities
Feature-preserving Discontinuity
Simple Example
Granular Flow
Gaussian FilteringRandom Walk Filtering
Particle Image Velocimetry
GaussianRandom Walk Particle Image Velocimetry
Landslide
Summary -Feature Preserving -Meshless -Interpretative -Flexible -Easy to implement
Limitations -Number of parameters -Dependency in them
Future Works - 3D vector field denoising algorithm
Thank you for your attention