07 Regionsegmentation Very Nice
07 Regionsegmentation Very Nice
07 Regionsegmentation Very Nice
Region Segmentation
Outline
Edge segmentation:
http://robots.stanford.edu/cs223b/index.html
Mean Shift: A Robust Approach toward Feature Space Analysis, by D. Comaniciu and P. Meer http://www.caip.rutgers.edu/~comanici/MSPAMI/msPamiResults.html
Region segmentation:
What is a region?
Region-based approach
Region-based segmentation
n i =1
Ri = I
Ri R j =
P ( Ri ) = TRUE ,
i = 1,2, n
i
P Ri R j = FALSE
3
2
3
5
2
3
4
5
4
5
6
3
6
6
7
9
12
11
11
8
7
3
10
11
9
12
10
8
4
22
15
17
18
18
3
2
9
10
19
16
9
3
1
3
3
1
2
5
1
threshold T10
3
2
3
5
2
3
4
5
4
5
6
3
6
6
7
9
12
11
11
8
7
3
10
11
9
12
10
8
4
22
15
17
18
18
3
2
9
10
19
16
9
3
1
3
3
1
2
5
1
3
2
3
5
2
3
4
5
4
5
6
3
6
6
7
9
12
11
11
8
7
3
10
11
9
12
10
8
4
22
15
17
18
18
3
2
9
10
19
16
9
3
1
3
3
1
2
5
1
3
2
3
5
2
3
4
threshold T11
5
4
5
6
3
6
6
7
9
12
11
11
8
7
3
10
11
9
12
10
8
4
22
15
17
18
18
3
2
9
10
19
16
9
3
1
3
3
1
2
5
1
3
2
3
5
2
3
4
threshold T12
5
4
5
6
3
6
6
7
9
12
11
11
8
7
3
10
11
9
12
10
8
4
22
15
17
18
18
3
2
9
10
19
16
9
3
1
3
3
1
2
5
1
region growing with variance of 2 in respect to value 11 with reference to threshold T11
3
2
3
5
2
3
4
5
4
5
6
3
6
6
7
9
12
11
11
8
7
3
10
11
9
12
10
8
4
22
15
17
18
18
3
2
9
10
19
16
9
3
1
3
3
1
2
5
1
3
2
3
5
2
3
4
5
4
5
6
3
6
6
7
9
12
11
11
8
7
3
10
11
9
12
10
8
4
22
15
17
18
18
3
2
9
10
19
16
9
3
1
3
3
1
2
5
1
3
2
3
5
2
3
4
5
4
5
6
3
6
6
7
9
12
11
11
8
7
3
10
11
9
12
10
8
4
22
15
17
18
18
3
2
9
10
19
16
9
3
1
3
3
1
2
5
1
3
2
3
5
2
3
4
5
4
5
6
3
6
6
7
9
12
11
11
8
7
3
10
11
9
12
10
8
4
22
15
17
18
18
3
2
9
10
19
16
9
3
1
3
3
1
2
5
1
seed
growing
nal region
http://uei.ensta.fr/baillie
The seed point can be selected either by a human or automatically by avoiding areas of high contrast (large gradient) => seed-based method.
Problem: To isolate the strongest lightning region of the image on the right hand side without splitting it apart.
Solution: To choose the points having the highest gray-scale value which is 255 as the seed points shown in the image immediately below.
threshold = 255
returns multiple
seeds
original image
threshold:
225~255
threshold:
190~225
threshold:
155~255
http://en.wikipedia.org/wiki/Region_growing
If the criterion of homogeneity is local (compared to the value of the candidate pixel and the pixel of the border) => linear method.
Threshold T:
P1 == P2 iff Diff(Col(P1),Col(P2)) < T
1
1 x
y
val=?
x==y: x<>y:
val = x
boundary(x)y if |x-y|T
new region index if |x-y|>T
1
1
2
1
2
1
3
1
http://uei.ensta.fr/baillie
http://uei.ensta.fr/baillie
original image
thresholding seg.
original image
split 1
split 2
split 3
http://uei.ensta.fr/baillie
RAG with adjacency relations (in red) for big black region.
RAG with adjacency relations (in red) for big black region.
http://astro.temple.edu/~siddu
watersheds
http://en.wikipedia.org/wiki/Watershed_(image_processing)#cite_note-1
Watershed segmentation
3D topographic surface
This technique aims at identifying all the third type of points (i.e., points of watershed lines) for segmentation
255
255
0 255
0 255
http://euclid.ii.metu.edu.tr/~ion528/demo/lectures/6/4/index.html
original image
gradient image
nal contours
http://cmm.ensmp.fr/~beucher/wtshed.html
original image
over-segmented image
http://cmm.ensmp.fr/~beucher/wtshed.html
http://en.wikipedia.org/wiki/Watershed_(image_processing)#cite_note-2
Snakes
Level sets
Graph cuts
Human
GPCA
Summary: