Image Segmentation

1
Image Segmentation
2
The Pixel Graph
Couplings Wij Reflect intensity
similarity
Low contrast strong coupling
High contrast weak coupling
3
Segmentationminimize cut coupling
Low-energy cut
4
Normalized-Cut Measure
Minimize
5
Coarsening the Minimization Problem
6
Coarsening Choosing a Coarse Grid
Representative subset
7
Importance of Soft Relations
8
Importance of Soft Relations
9
Importance of Soft Relations
10
Coarsening Choosing a Coarse Grid
Representative subset
11
Weighted Aggregation
aggregate
aggregate
12
Hierarchical Graph
13
Hierarchyin SWA
14
Segmentation by Weighted Aggregation
Detects the min-normalized-cut salient segments
Linear in of points (a few dozen operations per
point) Orders of magnitude faster than
15
Image Segmentation
16
Use Averages to Modify the Graph
17
Coarse-Scale Measurements

• Average intensities of aggregates

18
Image Segmentation
19
Recursive Measurements Intensity
intensity of pixel i
aggregate
average intensity of aggregate
20
Coarse-Scale Measurements

• Average intensities of aggregates
• Multiscale intensity-variances of aggregates

21
Isotropic Texture in SWA
Intensity Variance
Isotropic Texture of aggregate average of
variances in all scales
22
Coarse-Scale Measurements

• Average intensities of aggregates
• Multiscale intensity-variances of aggregates
• Multiscale shape-moments of sub-aggregates

23
Oriented Texture in SWA
Shape Moments

• center of mass
• width
• length
• orientation

Oriented Texture of aggregate orientation,
width and length in all scales
24
Previous (SWA)
New (SWA)
25
Mini Segments

• Keep the statistics of salient sub segments
  separately (all scales)
• Match aggregates by similarity of this statistics

26
Coarse-Scale Measurements

• Average intensities of aggregates
• Multiscale intensity-variances of aggregates
• Multiscale shape-moments of sub-aggregates
• Boundary alignment between aggregates

27
Boundary Integrity in SWA
28
Aggregate Measurements

• Average intensities

• Intensity variances (multiscale)

• Direction alignment

• Boundary alignment

• Average hair orientation

29
SWA
30
Specialized segmentationDetecting Lesions
Tagged
Our results
Data Filippi
31
Our Algorithm (SWA)
Ncuts
Isotropic texture SWA