1 - Geoscience & Remote Sensing Society

Report
Texture Segmentation for Remote Sensing Image
Based on Texture-Topic Model
Hao Feng
Zhiguo Jiang
Image Processing Center
Beijing University of Aeronautics & Astronautics
Xingmin Han
Beijing University of Technology
IGARSS 2011
water
sand
grass
tree 1, high density
tree 2, middle density
tree 3, low density
Proposed Method
-Topic Model: Latent Dirichlet Allocation
-LDA is a generative probabilistic model of
a corpus.
-LDA automatically clusters words into “topics”
and documents into mixtures of topics.
-Bag-of-Words Assumption
- Connecting word and feature descriptor
-Texture is topic,
pixel (feature descriptor) is word.
Previous Works
• Li Fei-Fei, Pietro Perona, CVPR 2005
• Supervised LDA
• Natural Scene Categorization
•Erik B. Sudderth, IJCV 2008
• Transformed Dirichlet Process
• Model natural scene with spatial constraint
•Marie Liénou,…, IEEE Geoscience and Remote Sensing Letter 2010
Dragos Bratasanu, Lon Nedelcu, Mihai Datcu, IGARSS 2011
•Annotation of Satellite Images Using LDA
•Xian Sun,…, IEEE Geoscience and Remote Sensing Letter 2010
• Model geospatial object using LDA
Latent Dirichlet Allocation
-LDA is a generative probabilistic model of a corpus.
-Documents are represented as random mixtures over latent topics
-where a topic is characterized by a distribution over words.
• Let’s assume that all the words within a
document are exchangeable.
Latent Dirichlet Allocation
N
p ( , z ,  |  ,  )  p ( |  )

p ( z n  ) p ( n z n ,  )
n 1
For each document,
• Choose  ~ Dirichlet()
• For each of the N words  n :
– Choose a topic zn ~ Multinomial()
– Choose a word  n from p ( n z n ,  ) , a multinomial probability
conditioned on the topic zn.
[blei 2003]
Latent Dirichlet Allocation
Topic: Education
labor
debt
……..
environment
undergraduace
postgraduate
course
student
education
University
Frequency
……..
Dictionary
word
This will mean that the Open University, which provides degree courses by distance
learning, will have among the lowest fees in England. Vice chancellor Martin Bean
promised "high-quality, flexible and great value-for-money education for all". The majority
of universities will charge £9,000 for some or all courses. More than two-thirds of the
Open University's students are studying part-time - and the university will be expecting to
benefit from the introduction of loans for part-time students. For a typical part-time Open
University student, studying at the level of half of full-time, the fees will be £2,500 per
year. Mr Bean said that the extension of the loan system represented the "beginning of a
new era for part-time students". Younger students At present the university has 264,000
students taking more than 600 undergraduate and postgraduate courses and professional
qualifications - …….
[BBC News]
Latent Dirichlet Allocation
θ
Topic Distribution
Building 2
Building 1
z
Latent topic
Topic 1
w
Bag-of-words
Topic 2
Topic 3
Spatial Constraint LDA
The William Randolph Hearst Foundation will give $1.25 million to Lincoln Center,
Metropolitan Opera Co., New York Philharmonic and Juilliard School. “Our board felt that
we had a real opportunity to make a mark on the future of the performing arts with these
grants an act every bit as important as our traditional areas of support in health, medical
research, education and the social services,” Hearst Foundation President Randolph A.
Hearst said Monday in announcing the grants. Lincoln Center’s share will be $200,000 for its
new building, which will house young artists and provide new public facilities. The
Metropolitan Opera Co. and New York Philharmonic will receive $400,000 each. The Juilliard
School, where music and the performing arts are taught, will get $250,000. The Hearst
Foundation, a leading supporter of the Lincoln Center Consolidated Corporate Fund, will
make its usual annual $100,000 donation, too.
2,600,000,000 results
448,000,000 results
13,400,000 results
57,100 results
Spatial Constraint LDA
Neighbors
Gaussian
Parameters
Spatial Constraint LDA
P ( n , rn , z n |  ,  , H ) 
P ( |  ) P ( z n |  ) P ( n | z n ,  ) P (  z ,  z |  ,  , ,  , z n ) P ( n |  z ,  z ,  n )
n
n
Normal Inverse Wishart
Dirichlet Distribution
n
n
Gaussian Distribution
Multinominal Distribution
Multinominal Distribution
1) For each image, Choose ~Dirichlet().
2) For each pixel, draw texture-topic zn ~ Multinominal() .
3) For a topic zn, choose Gaussian parameters
(  z n ,  z n ) ~   Wishart ( H )
4) Choose the visual word  n ~ Multino min al ( z n ,  )
5) Given the selected texture-topic zn and word  , choose word  n ~ Gaussian (  z ,  z ,  n )
n
n
n
Spatial Constraint LDA
z
w
Example:
Word
Red: Considered Word
(feature Descriptor)
r
Neighboring words
Experiments
1) Textures Segment
Brodatz texture and texture combination
4 dimension Haar feature
500 words visual dictionary
2) Remote Sensing Images
200 dimension DAISY descriptor
1000 words visual dictionary
Results
Texture model
Texture image
Visual word map
Texture image
Visual word map
Results
1
4
5
3
Region
1
Recall
0.94
False positive 0.06
2
0.92
0.02
3
0.91
0.01
4
0.89
0.02
5
0.99
0.09
2
1
4
5
3
1
Region
Recall
0.95
False positive 0.15
2
0.90
0.01
3
0.91
0.02
4
0.90
0.03
5
0.88
0.06
2
Results
Tree 1
Tree 2
Tree 3
Grassland
Road/Sand/Land
Garss
Road
Tree
Rooftop
Park
Conclusion
-Model Texture using LDA
-Introduce Neighborhood constraint to LDA
-Segment texture combinations and remote sensing
images
-Noise in sampling results
-Bag-of-words
-Speed
-Feature descriptor
-More information….
Thank you
[email protected]

similar documents