Beruflich Dokumente
Kultur Dokumente
;
IEEE Trans. on Im. Proc., 2009, 18:8, pp. 1830-1843
Early results
The Prague Segmentation Benchmark Application to the Berkley dataset
According to the different application goals, an image segmentation process can use visual information of many kinds
Intensities, colors, shapes, etc
Except for trivial cases, texture can have a deep heterogeneous visual appearence. An example from the remote sensing domain:
Except for trivial cases, texture can have a deep heterogeneous visual appearence An example from the remote sensing domain:
1. 2. 3.
Residential area
Vegetation/non-urban area
The texture of a surface represent its local visual appearance, where this locality has to be meant at multiple scales:
The texture of a surface represent its local visual appearance, where this locality has to be meant at multiple scales:
The texture of a surface represent its local visual appearance, where this locality has to be meant at multiple scales:
The texture of a surface represent its local visual appearance, where this locality has to be meant at multiple scales:
The texture of a surface represent its local visual appearance, where this locality has to be meant at multiple scales:
Key points
Description of visual appearance Focus on scales
Classical Approaches
Pixel-based analysis of low order neighbourhood Fixed scales
10
= Pr(+1 = | = , = )
(|)
11
The red and black color states present a strong interaction in some directions
12
Merging the red and black provides a 2-state representation. A hierarchy tree keeps track of the merging.
13
Two main issues for building the H-MMC model Retrieve elementary states (leaves) Study spatial interactions over different scales
15
= (n) , , =
P = P (n)
( |)
,,
L ,
, :
(L) , () ,
(L1) , (1) ,
(1) , (1) ,
16
Region Merging
17
Color map
Region Merging
CBC Block
Region Merging
CBC Block
SBC Block
Hierarchical map
Regions are sequentially merged to form the hierarchical structure and map
21
Hierarchical map
Regions are sequentially merged to form the hierarchical structure and map
22
Hierarchical map
Regions are sequentially merged to form the hierarchical structure and map
23
Hierarchical map
Regions are sequentially merged to form the hierarchical structure and map
24
25
The color segmentation provides color classes and uniform connected components (fragments) In the general case, fragments of the same color can belong to different textures:
At any scale, spatial interactions among different colors change For each color, fragments with similar contexts are clustered
Fragment Features
=1
, =
k
N k=1
[L 8]
k Pr(k )
26
The feature matrices are usually too big (but redundant) To ensure reliability to clustering, PCA reduction is used:
Weighted average over spatial directions 75% of energy preserved along the state dimension
27
Source Image
Red mask
Reduced Features
Clustering Result
28
In the last phase, regions are sequentially merged pairwise from the elementary states (leaves of the hierarchy tree) up to the root. The merging process is driven by a metric called Texture Score (TS) The Texture Score
()
(|)
At each step of the merging procedure, the state with lowest TS is merged to its dominant neighbor.
29
In the last phase, regions are sequentially merged pairwise from the elementary states (leaves of the hierarchy tree) up to the root.
log
= min log
() ( |)
log
To reinforce the contextual term, spatial closeness is enhanced using Kulback-Leibler divergence.
30
http://mosaic.utia.cas.cz
31
http://mosaic.utia.cas.cz
32
http://mosaic.utia.cas.cz
33
http://mosaic.utia.cas.cz
34
http://mosaic.utia.cas.cz
35
http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/segbench/
36
Panchromatic segmentation
PAN-MS Fusion
Region level processing
Spectral Clustering
To preserve fine details, only the panchromatic image is processed at pixel level and segmented. The resulting connected components are featured with a spectral signature and clustered to obtain the CBC map.
37
38
39
40
H-MMC is a new family of hierarchical image models aimed at describing the textural content, based on color and contextual properties. TFR is the corresponding hierarchical image segmentation algorithm, providing a set of nested segmentation maps.
Qualifying Points
Hierarchical segmentation Quick processing Unsupervised Broad-domain applications
Future Works
Morphological fragmentation at pixel level Geometrical descriptors for shape featuring Graph-based descriptors for spatial interactions New metrics for the texture merging phase
41