Transcript Diapositive 1

Inverse problems in
earth observation &
cartography
Inverse Problems in Earth
Observation and Cartography
Theme CogB
2
Project and People



Joint CNRS/INRIA/UNSA project team, created in 1998.
Project leader: Josiane Zerubia.
Members:
 3 INRIA: J. Zerubia (DR1), X. Descombes (CR1), I. Jermyn
(CR1).
 1 CNRS: L. Blanc-Féraud (DR2).
 1 postdoc.
 10 PhD students.
 5 Masters students (ENS Cachan, Sup’Aéro, ENAC, CPE
Lyon).
 5 student Interns (ETH Zurich, Szeged University , Sup’Com
Tunis, IIT New Delhi, IIT Roorkee).
 1 assistant INRIA (50%), 1 assistant CNRS/UNSA (10%).
3
Research domain
4
Principal goals
Taking into account the physics
of the sensor (visible, infrared,
radar, laser,…).
 Extraction of information relevant
for high-level interpretation.
 Reconstruction of 3D information
(digital elevation map) from 2D
data.
 Map updating for cartography.

5
Application areas

Commercial


Public interest


Precision farming, GIS,…
Forestry management,
environmental monitoring, urban
planning, cartography,…
Homeland security

Intelligence, pre- and post-mission
analysis,…
6
Scientific foundations

Probabilistic approaches



Variational approaches



MRFs: geometric properties, MRFs on trees,…
Stochastic geometry: marked point processes.
Regularization and functional analysis: BV space,
-convergence,…
Contours and regions: level sets, higher-order
active contours,…
Optimization/Parameter estimation


MCMC, RJMCMC,…
Diffusion processes.
7
Main contributions 2001-2005
8
Probabilistic models (I)

Markov Random Fields



Markov Random Fields for
hyperspectral data [IEEE TGRS
04]
New models on trees for
segmentation [IEEE TGRS 05]
Impact


Lyapunov Inst. (01-02); NATORussia (03-04); ECO-NET (05).
Contract and transfers: Alcatel
Alenia Space Cannes.
9
Probabilistic models (II)

Marked Point Processes



A new general modelling framework
[IEEE SP Mag. 02, IJCV 04, IEEE
PAMI 05, IJCV to appear].
New kernels for RJMCMC
optimization [LNS Springer 05].
Impact



PNTS (04-05); COLORS Arbres
(01); ARC MODE de VIE (05-06).
Contracts and transfers: DGA; IGN;
BRGM.
Eusipco04 Young Author Best
Paper award.
10
Variational methods (I)

Image decomposition [JMIV05]




New algorithm for decomposing an image
into geometry (in BV) + oscillations (in G).
Numerically challenging (L1 norm).
Applications: restoration, compression,
inpainting…
Impact



Math/STIC (02-03); ACI Multim (04-07).
Used for IR target detection by DGA.
Best PhD thesis prize 05 from EEA Club,
Signal and Image section.
11
Variational methods (II)

Higher-order active contours



New method for the inclusion of
geometric information in active
contour models. [IJCV to appear]
Reformulation as phase fields to
simplify implementation and allow
model learning. [ICCV05]
Impact



EU NoE MUSCLE (04-08); ACI
QuerySat (04-06).
Contract: Alcatel Alenia Space
Cannes.
Transfers: U. Szeged, Hungary;
LIAMA, China.
n(p)
n(p0)
p
p0
Long-range
interactions
Effect of prior
12
Parameter Estimation



Estimation of model hyperparameters
for deconvolution of HR visible and
TIR data.
Estimation of sensor parameters for
blind deconvolution of HR visible data.
Impact



Contracts and transfers: CNES; Astrium;
Sagem.
Patent #0110189 (France: 01; Europe,
Israel, USA, Canada, Japan: 02).
Usage rights granted to French Space
Agency for future satellites.
13
Application to other domains:
astrophysics

Image deconvolution


Galaxy filament detection


Using complex wavelet packets.
Using a marked point process.
Impact



COLORS DECONVOL (02);
COLORS FIGARO (05).
Transfers: Côte d’Azur
Observatory.
First automatic extraction of
galaxy filaments from a real
galaxy catalogue (provided by
Harvard).
14
Application to other domains:
confocal microscopy

Deconvolution


Using Richardson-Lucy algorithm taking
into account the edges of the object (TV
regularization) [MRT Journal to appear]
Impact


ARC DeMiTri (03-04); P2R Franco Israeli
Programme (05-06).
Transfers: Pasteur and Weizmann Institutes.
15
Project-team positioning
16
Positioning: INRIA scientific
challenges

Develop multimedia data and multimedia
information processing (40%):




Couple models and data to simulate and
control complex systems (15%)


EU project MOUMIR (00-04).
EU NoE MUSCLE (04-08).
ACI QuerySat (04-07)
ARC MODE de VIE (05-06)
Model living structures and mechanisms
(15%)


ARC DeMiTri (03-04)
P2R Franco-Israeli programme (05-06)
17
Positioning: INRIA (I)

Digiplante: peer.


Imedia: peer.



Strong collaboration on forestry growth modelling
(ARC Mode de Vie).
Joint partners in EU NoE MUSCLE & ACI QuerySat.
Strongly complementary roles: Imedia = database
retrieval; Ariana = image processing.
Vista: peer.


Joint partners in EU NoE MUSCLE.
Benchmarking of segmentation methods developed in
the two project-teams.
18
Positioning: INRIA (II)

Clime: peer.


Epidaure/Asclepios: peer.



Joint partners in EU IMAVIS project.
Complementary work on biological microscopy imaging but
otherwise different applications.
Mistis: peer/competitor.



Works on satellite data, but mainly for atmospheric & meteorological
modelling & on time series.
Collaboration with previous project-team IS2.
Methodological overlap (MRFs, variational, & comparison) but
different applications.
Odyssée: peer/competitor.



Joint partners in EU IMAVIS project.
Shape from shading: different approaches (stochastic vs. PDEs).
Variational methods (different applications) & shape information
(different approaches).
19
Positioning: France (I)

Signal & Image Processing department, ENST,
Paris: peer/competitor. (Campedel, Maître,
Nicolas, Roux, Sigelle, Tupin)



Main competitor in France for remote sensing
applications.
Joint partners in ACI QuerySat & in PNTS initiative;
joint PhD.
PASEO group, MIV team, LSIIT, Strasburg: peer.
(Collet, Jalobeanu)

New group: possible collaboration on remote sensing.
20
Positioning: France (II)

CMLA, ENS Cachan: peer. (Aujol, Chalmond, Morel,
Younes)



CMAP, École Polytechnique, Palaiseau: peer.
(Chambolle)


Joint partners in ACI Multim & MATH/STIC project.
Complementary work on variational and stochastic methods.
Joint partners in ACI Multim & joint work on functional analysis for
image processing.
CEREMADE/Paris XIII: peer/competitor. (Cohen, Dibos)


Variational methods for shape description: different approaches
but some overlap of application domains.
Complementary work on TV regularization and PDEs.
21
Positioning: international (I)

Stochastic geometry:




Ariana is one of the key developers of object
processes for image analysis and understanding.
Collaboration with CWI (van Lieshout). Peer.
Related to work at Brown (Grenander), Florida State
(Srivastava). Simpler objects but complex inter-object
interactions. Peer.
German Space Agency DLR and U. Jaume I have
adapted models developed by Ariana.
Peer/competitor.
22
Positioning: international (II)

Higher-order active contours:


Distinct from other methods for the inclusion of prior
geometric information.
Better adapted to remote sensing than the many
template-based methods:

U. Bonn (Cremers); U. Florida (Chen, Thiruvenkadam, Huang,
Wilson, Geiser); Yale University (Tagare); U. Mannheim
(Kohlberger, Schnorr); UCLA (Soatto); Saarland University
(Weickert); Siemens Corporate Research (Paragios); MIT
(Leventon, Grimson, Faugeras);… Peers/competitors.
23
Positioning: international (III)

Functional analysis for image processing:




Image Processing group at UCLA (Chan,
Esedoglu, Osher, Vese). Also work on functional
analytic and PDE approaches. Peer/competitor.
U. Pompeu Fabra Barcelona (Caselles). Related
work on image inpainting, and strong links via
exchange of students and researchers. Peer.
U. Minnesota (Sapiro). Works more on PDEs than
variational methods. Peer/competitor.
CNR Rome (March). Collaboration on convergence. Peer.
24
Evolution of the objectives
during 2001-2005

All the objectives defined in 2001 have been
reached except:

The work on parameter estimation for blind
deconvolution of optical and infrared satellite
images was stopped due to legal problems over
industrial property between INRIA-CNRS vs CNES
or Sagem).


Despite patent deposit in 2001 and extension in 2002.
Two new topics emerged:


Astrophysical image processing.
Biological image restoration and deconvolution.
25
Recommendations from
previous evaluation

Ariana has continued to:




Develop innovative theoretical methods
Publish in the best international journals and conferences.
Maintain high international visibility.
Ariana has increased its activity in:

Validation:



Applications:


Via data and ground truth obtained from end users (CNES,
IAURIF, IGN, IFN, BRGM, Alcatel Alenia Space).
Via end user validation and use (IGN, BRGM).
Environmental research (IFN, Alcatel Alenia Space, Silogic):
trees, fires,…
Image retrieval:

Collaborations with INRIA project-teams, French and European
researchers (ACI QuerySat, EU projects MOUMIR and
MUSCLE, Math/STIC Tunisia) and end users (CNES, DLR).
26
Goals for the next four years
27
Future: probabilistic methods

Multi-object processes.



Different types of objects (e.g. roads, buildings,
trees,…) will be included in a single model in which
they mutually interact.
High-risk.
Improved models of texture.


Quartic models of the joint statistics of adaptive
wavelet packet coefficients.
Medium-risk.
28
Future: variational methods

Image decomposition & restoration.



Detection of objects of codimension > 1 in 3D images.



Multispectral extension using inter-band correlation.
Medium-risk.
New functionals based on Ginzburg-Landau theory.
High-risk.
Higher-order active contours & phase fields.


Multiscale/wavelet models.
Medium- to high-risk.
29
Future: parameter estimation

Parameter estimation for marked point processes
using new diffusion processes & comparison to
RJMCMC.


Model learning for higher-order active contours &
phase field models.


High-risk.
Parameter learning and resolution dependence for
adaptive wavelet packet texture models.


Medium- to high-risk.
Medium-risk.
Estimation of image acquisition parameters for the
blind deconvolution of microscopy images.

High-risk.
30
Future: applications

Automatic urban scene analysis & cartography.




Forest monitoring (tree population counting &
classification, disaster damage evaluation).



Using multi-object marked point processes.
Using phase field higher-order active contours.
High-risk.
Using marked point processes & Markov random fields.
Medium-risk.
Blind deconvolution of microscopy imagery.


Using physical models of the image acquisition process and
TV and wavelet regularizations.
Medium- to high-risk.
31
Some statistics 2001-2005

Publications:






PhDs & Habilitations: 13.
Books edited: 2.
Journal articles + book
chapters: 44.
Conference articles: 80.
Research reports: 49.
Contracts:


Industrial: 15.
Academic: 29.

Software:






Transfers: 13.
Patents: 2.
Student interns: 23.
Visiting scientists: 62.
Seminars: ~100.
Teaching: ~550h taught.