150 likes | 264 Views
Shape Priors and Knowledge Based Segmentation. By Neva Waynesboro. Background/Introduction. The goal of segmentation is to simplify and/or change the representation of an image into something that is more meaningful and easier to analyze. Result. Input. Overview.
E N D
Shape Priors and Knowledge Based Segmentation By Neva Waynesboro
Background/Introduction • The goal of segmentation is to simplify and/or change the representation of an image into something that is more meaningful and easier to analyze Result Input
Overview Image Segmentation - Labeling Problem • Find a mapping from a set of sites S to a set of labels L. • The set of sites S corresponds to the set of pixels • The set of labels L corresponds to the set of objects we want to segment L = {0,1} : Foreground/Background Segmentation
Over Cont. Gibbs Distribution : measures the cost of assigning a given set of labels to all the sites in the clique c. Clique Potential
The Problem Database of Heart Annotations Result Input Image ? • Proposed Solution: • Shape-constrained contour evolution formulated as a MAP-MRF Problem
Model Description Labels of the MRF = local displacements of the contour points Learn shape by learning the distance between contour points Sites of the MRF = points on the contour (explicit contour representation) Ahmed Besbes, N. K., Georg Langs, Nikos Paragios (2009). "Shape Priors and Discrete MRFs for Knowledge-based Segmentation." IEEE: 1295 1302.
Proposed Solution • Image Segmentation • We cast the segmentation problem as a MAP-MRF problem [1]. We compute the MAP-MRF solution by minimizing the following Gibbs energy function: • We represent the segmentation energy as a combination of two energy functions, each modeling a specific type of prior information, as shown below:
Mumford shah energy Divergence theorem Appearance Prior
Appearance Prior Cont. We then set and get Now
Shape Prior Models prior information about the shape of the object being segmented by, Learn shape by learning the distance between contour points
Steps • Segmentation Algorithm • Repeat • Step1: Compute Appearance Prior • Step 2: Compute Shape Prior • Step 3: Minimize MRF energy using belief propagation • until convergence is reached
Results Input Image Initial Contour with Normals Appearance Shape Result The problem was that the appearance was not growing correctly, so when applied both to the result just moved rather than actually grow.
References • Ahmed Besbes, N. K., Nikos Paragios (2009). "Graph-Bases Knowledge-Driven Discrete Segementation Of The Left Ventricle." IEEE: 49 - 52. • Ahmed Besbes, N. K., Georg Langs, Nikos Paragios (2009). "Shape Priors and Discrete MRFs for Knowledge-based Segmentation." IEEE: 1295 1302. • D.R. Chittajallu , E. S., O.C. Avila-Monets, R.P. Yalamanchili MRF-Based Solutions to Image Analysis Problems: An Investigation. Houston University of Houston.