segmentation-summary and evaluation techniquesaalbu/computer vision 2009/lecture 14...1...

25
1 Segmentation-summary and evaluation techniques Section 6.5 and Chalana and Kim (1997) “A Methodology for Evaluation of Boundary Detection Algorithms on Medical Images”

Upload: others

Post on 10-Aug-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

  • 1

    Segmentation-summary and evaluation techniques

    Section 6.5

    and

    Chalana and Kim (1997)“A Methodology for Evaluation of Boundary Detection Algorithms on Medical Images”

  • 2

    Segmentation - summary

    ThresholdingEdge-basedRegion-basedModel-based

    HoughTemplate matching (supervised)

  • 3

    Thresholding

    Simplest segmentation processFocus is on automatic thresholding

    Histogram-basedStatistical properties ( 1st order K means, 2nd-order Otsu)

    Pros:FastWorks well for detecting non-touching objects of constant reflectivity or light absorption of their surfaces

    Cons:segmentation results may contain holes inside regions corresponding to objects or several objects merged into one region

  • 4

    Edge-based

    Work with results of edge detectors (discontinuities in gray-level, colour, texture etc)Main goal: grouping local edges into an image where only edge chains with a correspondence to existing objects or image parts are presentCommon problems:

    Sensitivity to image noiseLow contrasted borders

  • 5

    Region-based

    Image is partitioned into a minimal number of regions that satisfy an homogeneity criterionCommon problems

    Choice of homogeneity criterion: if too low, undersegmentation; if too high, oversegmentationPixel aggregation: choice of seed pixel; choice of similarity measure; choice of similarity thresholdResults of region merging are typically undersegmentedResults of region splitting are typically oversegmented and have a blocky aspect Split-and-merge approaches

  • 6

    Model-based segmentation

    Hough transformBased on a geometric model of the primitive; model is parametricDetection of the primitive is equivalent to detection of maxima in the accumulator space. Maxima detection in accumulator space is typically performed by thresholdingFast for lines, reasonably fast for circles, computational complexity increases exponentially with number of parameters in the geometric description of the primitiveLimited applicability

  • 7

    Model-based segmentation

    Template matchingSupervised segmentation (we know what we are looking for)

    For fixed size templates: Image is scanned and a similarity measure (correlation) is performed between the neighborhood centered at each image location and the templateIf target objects have variable size, image pyramids (gaussian smoothing of various levels of detail)Problems: does not work for deformable objects, is not view-invariant

  • 8

    More advanced segmentation techniques

    Graph-cut segmentationDeformable models

    SnakesLevel Sets

  • 9

    Region segmentation: minimum-cost graph cutA graph cut is a partitioning of a graph where the cost of the cut is thecumulative cost of cutting the respective arcs (links).

  • 10

    Deformable models

    Powerful top-down techniqueUses a model (prior knowledge) and deforms it to fit the image data (current information)Main idea:

    1. Establish the problem as the minimization of some cost function.2. Use established optimization techniques to find the optimal

  • 11

    Active contours (snakes)

    Based on the minimization of an energy functionThe original formulation of this “energy” was

    Esnake = wintEint + wimageEimage + wconEconwhere each term is as follows:

    Eint Internal Energy (Keeps the snake from bending too much)Eimage Image Energy (Guides the snake along important image features)Econ Constraint Energy (Pushes or pulls the snake away from or towards user-defined positions)

    The total energy for the snake is the integral of the energy at each point:

  • 12

    Level sets

    Implicit formulation of the curvesSegmentation not affected by topological changesAn unknown number of objects can be detected simultaneously

  • 13

    Evaluating segmentation techniques

    As in other areas of vision, evaluation is a problemWe need to know what the correct result isWe need some way to compare the result of each algorithm to the ideal situation

    From Tony Pridmore’s Lecture Notes on Image Processing and Interpretation, University of Nottingham

  • 14

    Evaluating segmentation

    Possible approachesGround truth – get a ‘correct’segmentation and compare the results of the algorithm to itEvaluations based on region properties – we want the regions to be uniform, and for adjacent regions to be differentEvaluating robustness

    If we deliberately introduce noise or partially mask the object of interest, how will the segmentation result be affected?

    Adapted from Tony Pridmore’s Lecture Notes on Image Processing and Interpretation, University of Nottingham

  • 15

    Ground truth segmentation

    Typically used in medical imaging applicationsIssue: human segmentations can vary significantlyHow do we build a ground truth segmentation from several human segmentations?

  • 16

    Statistical ground truth

  • 17

    Additional reading (ELEC 536)

    Chalana And Kim: Evaluation Of Boundary Detection Algorithms On Medical Images, IEEE Transactions on Medical Imaging, 1997

  • 18

    Ground truth in other applications

    Experiment: segmenting an image by hand

    Adapted from Tony Pridmore’s Lecture Notes on Image Processing and Interpretation, University of Nottingham

  • 19

    Ground truth in other applications

    Experiment: segmenting an image by hand

    Adapted from Tony Pridmore’s Lecture Notes on Image Processing and Interpretation, University of Nottingham

  • 20

    Ground truth in other applications

    Human segmentation of complex scenes is subjective; it depends on visual representation among many other things Are human segmentations consistent?

    Adapted from Tony Pridmore’s Lecture Notes on Image Processing and Interpretation, University of Nottingham

  • 21

    Comparing image segmentations

    Suppose we have a agreed ground truthWe need to compare two sets of regionsWhat does it mean for two sets of regions to be similar?Is the number of regions important?Does it matter if two regions are merged or if one is split in two?

    Ground truth partition Which result is better?

    Adapted from Tony Pridmore’s Lecture Notes on Image Processing and Interpretation, University of Nottingham

  • 22

    Current measures of similarity

    Applicable when only one region of interest in imageRegion-based: Mutual overlap (6.5.1.1)Limits

    Does not give any information about boundariesConceals quality differences between segmentationsAssumes a closed contourLarge errors for small objects

  • 23

    Current measures of similarity

    Border-based (6.5.1.2)Hausdorff distance

    Idea: consider pixels in the two contours as sets of finite points

  • 24

    Hausdorff distance (cont’d)

    http://en.wikipedia.org/wiki/File:Hausdorff_distance_sample.svg

  • 25

    Unsupervised evaluation

    We can compute statistics about theregions

    The averages of adjacent regions should be differentThe standard deviation within a region should be low

    Segmentation-summary and evaluation techniquesSegmentation - summaryThresholdingEdge-basedRegion-basedModel-based segmentationModel-based segmentation More advanced segmentation techniquesSlide Number 9Deformable modelsActive contours (snakes)Level setsEvaluating segmentation techniquesEvaluating segmentationGround truth segmentationStatistical ground truthAdditional reading (ELEC 536)Ground truth in other applicationsGround truth in other applicationsGround truth in other applicationsComparing image segmentationsCurrent measures of similarityCurrent measures of similarityHausdorff distance (cont’d)Unsupervised evaluation