13.07.2015 Views

Graph based image segmentation with sub- pixel precision

Graph based image segmentation with sub- pixel precision

Graph based image segmentation with sub- pixel precision

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

<strong>Graph</strong> <strong>based</strong> <strong>image</strong><strong>segmentation</strong> <strong>with</strong> <strong>sub</strong><strong>pixel</strong><strong>precision</strong>Filip MalmbergCentre for Image analysisSwedish University of Agricultural SciencesUppsala University


Image <strong>segmentation</strong> and feature2●●●measurementsImage analysis: Measure properties of objects,<strong>based</strong> on a sampled <strong>image</strong>.Usually requires <strong>segmentation</strong>.Measurement accuracy limited by <strong>image</strong>resolution.namn@cb.uu.se


3Crisp/Fuzzy <strong>segmentation</strong>●●A crisp <strong>segmentation</strong> maps each <strong>image</strong>element to exactly one object category.A fuzzy <strong>segmentation</strong> allows an <strong>image</strong> elementto partially belong to many different objectcategories.namn@cb.uu.se


4Pixel coverage <strong>segmentation</strong>●●Special case of fuzzy <strong>segmentation</strong>Let the degree of membership to a certain classbe proportional to the extent to which the <strong>image</strong>element is covered by that object.namn@cb.uu.se


5Pixel coverage <strong>segmentation</strong>-How to digitize?vs.namn@cb.uu.se


6Pixel coverage <strong>segmentation</strong>●●By using feature estimators that work directlyon this type of representation, we can get muchmore precise feature measurements!To benefit from this in practice, we needsuitable <strong>segmentation</strong> methods.namn@cb.uu.se


7<strong>Graph</strong> <strong>based</strong> <strong>segmentation</strong>●●Interpret <strong>image</strong> as an edge weighted graph.Two common ways to represent a<strong>segmentation</strong>: <strong>Graph</strong> cuts and graph labelingnamn@cb.uu.se


Pixel coverage <strong>segmentation</strong> on8●●graphsHow do we define ”<strong>pixel</strong> coverage” on graphs?No notion of the ”shape” of a <strong>pixel</strong>namn@cb.uu.se


9General idea●Interpret edges as paths between vertices.● We can parametrize a path by t ∈ [0,1].●Define <strong>segmentation</strong> labels for all points alongthe edges – Edge <strong>segmentation</strong>.namn@cb.uu.se


10Located cuts●Define a point along each edge in the cut wherethe transition occurs.namn@cb.uu.se


11Vertex coverage <strong>segmentation</strong>●●Integrate labels over the domain of each vertex.The domain (”shape of a vertex”) is the set of”half-edges” adjacent to the vertex:●For an edge <strong>segmentation</strong> generated by alocated cut, we get closed formulas that areeasy evaluate.namn@cb.uu.se


12Vertex coverage <strong>segmentation</strong>namn@cb.uu.se


13Computing located cuts in practice●Two methods proposed:▬▬Sub-<strong>pixel</strong> IFTSub-<strong>pixel</strong> ”defuzzification”namn@cb.uu.se


14Does it make a difference?namn@cb.uu.se


15Yes!namn@cb.uu.se


16Synthetic objects● Measurements were made on 2000 objects.namn@cb.uu.se


17Interactive <strong>segmentation</strong>●Study variation of area measurement <strong>with</strong>respect to changes in seed-point placement.namn@cb.uu.se


18Results●Both experiments show that using the proposedframework leads to significantly better stabilitywhen measuring object features.namn@cb.uu.se


19●Results, exampleStandard deviation of relative measurementerror reduced by a factor 5.namn@cb.uu.se


20Ideas for future work●●Use located cuts as a general objectrepresentation on graphs?Can we perform operations (e.g. Morphology)directly on this representation?namn@cb.uu.se


21References●●●Sub-<strong>pixel</strong> Segmentation <strong>with</strong> the Image Foresting Transform,F. Malmberg, J. Lindblad, I. Nyström, In Proceedings of the 13th International Workshop onCombinatorial Image Analysis 2009 (IWCIA'09)Image Foresting Transform: On-the-fly Computation of RegionBoundaries, F. Malmberg, In Proceedings of Swedish symposium on <strong>image</strong> analysis 2010(SSBA'10).A graph <strong>based</strong> framework for Sub-<strong>pixel</strong> Image Segmentation,F. Malmberg, J. Lindblad, N. Sladoje, I. Nyström. Submitted for publication.namn@cb.uu.se

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!