Automatic Target Recognition in high resolution Optical Aerial Images

Page 17th CNES/DLR Workshop on Information Extraction and Scene Understanding for Meter Resolution Image – 29/03/07 - Oberpfaffenhofen

EADS DS / SDCLTIS

Automatic Target Recognition in high resolution Optical Aerial Images

7th CNES/DLR Workshop on Information Extraction and Scene Understanding for Meter Resolution Image

Xavier PERROTTON Marc STURZEL Michel ROUX [email protected] [email protected] [email protected]

Image & Signal Processing Laboratory Telecom Paris

Page 2

EADS DS / SDCLTIS

7th CNES/DLR Workshop on Information Extraction and Scene Understanding for Meter Resolution Image – 29/03/07 - Oberpfaffenhofen

Objective :

make a breakthrough on ATR in visible images Context

– Observation systems (satellites, UAVs, aircraft…)

· Huge volume of data sent back by current and future systems

· Limited number of operators

· Pressure to shorten the loops

– Autonomous systems (missiles, UAVs…)

· More intelligence onboard

Strong need in the future for :

– Fully automatic processing

– Autonomous systems

ATR still unsolved for operational use

Why ATR for EADS?

Page 3

EADS DS / SDCLTIS


The problem

• Challenging problems• Lighting, occlusion and background • Difficult segmentation• Targets size

• Local descriptors approach• Learning appearance characteristics • Focusing on discriminative parts of the target

Page 4

EADS DS / SDCLTIS


Questions

Can we efficiently use local descriptors?

How to extend application domain by statistical learning?

Page 5

EADS DS / SDCLTIS


Local descriptors: method

2 Generating a list of candidate matches3 Defining an hypothesis

4 Hypothesis propagation

Recognized target

1 Selecting and learning keypoints

Descriptor : GLOH (Gradient Location orientation Histogram)

Page 6

EADS DS / SDCLTIS


Local descriptors: method

Descriptors

· GLOH (Gradient Location

orientation Histogram) [1]

How to match Keypoints?

· Looking for the best

match on each pixel

· Associating a limited

number of matched points

for each learned keypoint

How to define an hypothesis ?

· Choosing three points

among the best matches

· Evaluating the affine

transform

How to propagate an hypothesis?

· Checking for agreement

between each candidate

point and the geometric

model

[1] K. Mikolajczykand C. Schmid. A performance evaluation of local descriptors. In Proc. IEEE CVPR, June 2003

Page 7

EADS DS / SDCLTIS


Local descriptors: tests on real images (1)

Learned target Matched targets

Page 8

EADS DS / SDCLTIS



Learned target

Matched targets

Page 9

EADS DS / SDCLTIS



XAerial Images difficulties :

• Few points• Not robust to background

We must find a way to learn the variability of appearance characteristics

Page 10

EADS DS / SDCLTIS


AdaBoost: a powerful learning concept

Principle :

– Iterative learning algorithm introduced by Freund and Schapire [2]

– Constructing a “strong” classifier in combining “weak” classifiers

– Selecting a “weak” classifier at each iteration

Used for face detection by Viola and Jones [3] Advantages :

– often outperforms most “monolithic” strong classifiers such as Neural

Networks

– Few parameters to tune[2] Y. Freund and R. E. Schapire. A decision-theoretic generalization of on-line learning and an application to boosting. 97

[3] Paul Viola and Michael J. Jones. Rapid Object Detection using a Boosted Cascade of Simple Features.IEEE CVPR, 2001

Page 11

EADS DS / SDCLTIS


AdaBoost: algorithm

Adaboost starts with a uniform distribution of

“weights” over training samples

We obtain a weak classifier from the weak

learning algorithm, hj(x) at each round

We compute j that measures the confidence

assigned to hj(x)

We increase the weights on the training

samples that were misclassified

Repeat

At the end, make a weighted linear

combination of the weak classifiers obtained

at all iterations

)()()( 11final xxx nnhhf

Page 12

EADS DS / SDCLTIS


Weak classifier

Feature

X

X

X

X

XX

X

XX

X

XX

X

X

X

X

XX

Feature output databaseDatabase

Positive, negative samples

Feature + Threshold =

weak classifier

A weak classifier is only required to be better than chance Very simple and computationally inexpensive

• Haar like features

• Gabor filters

• Steerable filters

• orientation estimation features…

Page 13

EADS DS / SDCLTIS


Database

Generation of a representative database with positive and negative samples

The classifier is learned on images of fixed size

Detection is done through a sliding search window

Angle variations : -5° to 5°

Page 14

EADS DS / SDCLTIS


Tests on real images

Learned different appearance characteristics successfully

Page 15

EADS DS / SDCLTIS


Descriptors

Challenge :

Finding descriptors less sensitive to background and target texture

Haar like features learn only difference of contrasts

– Not enough to discriminate complex textures

– But can be very efficient on shadow Gabor filters, steerable filters, orientation estimation features

– More robust to background and target texture

Page 16

EADS DS / SDCLTIS


Conclusion

Local descriptors enable to define an efficient ATR algorithm

– Targets can be modelled as a collection of regions

– Geometric constraints are efficient to eliminate false alarms

Statistical learning enables to extend the application domain

– Selecting the discriminating features

– Learning the variability of appearance characteristics

– Descriptors

- To detect particular oriented edges

- To detect different regions

Automatic Target Recognition in high resolution Optical Aerial Images

Documents

pagelocal descriptors

weak classifiers

weak learning algorithm

real images

negative samplesthe

images of f

statistical learning

iterative learning algorithm