Image Segmentation with Graph Cut

Document Sample
Image Segmentation with Graph Cut Powered By Docstoc
					Image Segmentation
                with
       Graph Cut

               Wei Wu
          Mentor: Nhat Vu
  Faculty Advisor: B. S. Manjunath
        Vision Research Lab
       Image Segmentation
• Computer algorithm divides
  image into meaningful parts
• Uses: quantitative analysis,
  content-based image
  search
• Computer-automated
  image segmentation is
  faster & more reproducible
  than human calculation
• My method of image
  segmentation: GRAPH CUT
Graph Cut: A Real Life Application
                 • Problem: Which distributor should
                   deliver to each store?
                 • Assign cost of going from one store
                   to its neighbor (black links) based
                   on efficiency of route
                    – Efficiency defined by traffic, road
                      quality, etc.
                 • Solutions:
                     Arbitrarily divide stores between the
                      two distributors
   Starbucks
   Distributor       Try every possible division to find most
                      efficient
                    Use “graph cut” algorithm to
                      eliminate least efficient route
                      (Boykov & Kolmogorov 2004)
 Graph Cut Applied to Images


  2x3         node for nodes linked by   edges given   algorithm
image        each pixel    edges            costs      makes min.
                                                        cost cut
• My project: Assign costs to edges
• Edge cost is based on degree of similarity between
  connected pixels
   – Higher similarity between neighbors means higher cost
     of cutting the edge between them
• What defines similarity?
   – Intensity (brightness)       – Texture
 Defining Costs w/ Feature Space



                         (226, 5, 10)



 3x5                     plot features
image                     into multi-
        3 channels → 3   dimensional
        features            space
   Defining Costs w/ Feature Space

                                           bigger feature
                                              distance


                                             smaller cost
                                               (vice versa)



                       k = some constant
Cost Function: e-k*d   d = feature distance btwn. 2 neighbor pixels
Intensity-based Segmentation
Color-based Segmentation


 red channel

       +
 green channel

       =
 both channels
 Training & Feature Weighting
Use prior knowledge about regions to train for which
features are better for differentiating the regions.

            # pixels
                           B   A     small overlap
                                       region
            # pixels




                           A   B   feature good for
                                   differentiating the
                                       two regions
      B
            # pixels




  A                                    (vice versa)
                       A       B
Segmentation w/ Weighted
       Features
 train on R & G




                  A   B


image


 train on B & G
                      A   B
               Conclusion
• For more complex
  images like retinal
  images, more features
  must be used for
  segmentation
• Graph Cut has potential
  to be useful for every-
  day applications
  – Image cropping
  – Colorization
 Acknowledgements

            Nhat Vu

 Professor Manjunath & the VRL

       Jens, Mike, Evelyn

             CNSI

My fellow Apprentice Researchers