SAR%20Automatic%20Target%20Recognition%20Proposal

1
SAR Automatic Target Recognition Proposal

• J.Bell, Y. Petillot

2
Contents

• Background
• ATR on SAR
• ATR on Sonar
• Supporting Technologies
• Initial results on SAR
• Way forward

3
ATR approaches
4
Unsupervised Techniques

• Future automated systems will require all
  available information (navigation data, image
  processing models .etc.) to be fused.

5
CAD/CAC Proposal
REMOVE FALSE ALARM
1
2
YES
Detect MLOs (MRF-based Model)
Extract Highlight/Shadow (CSS Model)
False Alarm?
NO
Fuse Other Views
Classify Object (Dempster-Shafer)
Positive Classification?
MINE
YES
NO
6
The Sonar Process

• Sonar images represent the time of flight of the
  sound rather than distance.
• Objects appear as a highlight/shadow pair in the
  sonar image.

7
The Detection Model

• A Markov Random Field(MRF) model framework is
  used.
• MRF models operate well on noisy images.
• A priori information can be easily incorporated.

• They are used to
• retrieve the underlying label field (e.g
  shadow/non-shadow)

8
Basic MRF Theory

• A pixels class is determined by 2 terms
• The probability of being drawn from each classes
  distribution.
• The classes of its neighbouring pixels.

9
Incorporating A Priori Info

• Object-highlight regions appear as small, dense
  clusters.
• Most highlight regions have an accompanying
  shadow region.

Segment by minimising
10
Initial Detection Results
DETECTED OBJECT

• Initial Results Good.
• Model sometimes detects false alarms due to
  clutter such as the surface return requires
  more analysis!

11
Object Feature Extraction

• The objects shadow is often extracted for
  classification.
• The shadow region is generally more reliable than
  the objects highlight region for classification.
• Most shadow extraction models operate well on
  flat seafloors but give poor results on complex
  seafloors.

12
The CSS Model

• 2 Statistical Snakes segment the mugshot image
  into 3 regions object-highlight, object-shadow
  and background.

• A priori information is modelled
• The highlight is brighter than the shadow
• An objects shadow region can only be as wide as
  its highlight region.

13
CSS Results
Standard Model
CSS Model
14
The Combined Model

• Objects detected by MRF model are put through the
  CSS model.
• The CSS snakes are initialised using the label
  field from the detection result. This ensures a
  confident initialisation each time.
• The CSS can detect MANY of the false alarms.
  False alarms without 3 distinct regions ensure
  the snakes rapidly expand, identifying the
  detection as a false alarm.
• Navigation info is also used to produce height
  information which can also remove false alarms.

15
Results
16
Results 2
17
Results 3
18
Result 4
19
BP 02 Results

• The combined detection/CSS model was run on 200
  BP02 data files containing 70 objects.
• 80 of the objects where detected and features
  extracted(for classification).
• 0.275 false alarms per image.
• The surface return resulted in some of the
  objects not being detected. Dealing with this
  would produce a detection rate of 91.

20
Object Classification

• The extracted objects shadow can be used for
  classification.
• We extend the classic mine/not-mine
  classification to provide shape and dimension
  information.
• The non-linear nature of the shadow-forming
  process ensures finding relevant invariant
  features is difficult.

Shadows from the same object
21
Modelling the Sonar Process

• Mines can be approximated as simple shapes
  cylinders, spheres and truncated cones.
• Using Nav data to slant-range correct, we can
  generate synthetic shadows under the same sonar
  conditions as the object was detected.
• Simple line-of-sight sonar simulator. Very fast.

22
Comparing the Shadows

• Iterative Technique is required to find best fit.
  Parameter space limited by considering highlight
  and shadow length.
• Synthetic and real shadow compared using the
  Hausdorff Distance.
• It measures the mismatch of the 2 shapes.

HAUSDORFF DISTANCE
23
Incorporating Knowledge

• As the technique is model-based, information on
  likely mine dimensions can be incorporated.
• Limited information from the highlight region can
  also be used to distinguish between the tested
  classes.
• We obtain an overall membership function for each
  class.

24
The Classification Decision

• A decision could be made by simply defining a
  Positive Classification Threshold. This is a
  hard decision and non-changeable.
• The lawnmower nature of Sidescan surveys
  ensures the same object is often viewed multiple
  times. The model should ideally be capable of
  multi-view classification.
• We use DEMPSTER-SHAFER theory.

25
Mono-view Results

• Dempster-Shafer allocates a BELIEF to each class.
• Unlike Bayesian or Fuzzy methods, D-S theory can
  also consider union of classes.

26
Mono-view Results
Model was tested on 66 mugshots containing
cylinders, Spheres, Truncated cones and clutter
objects.
27
Multi-view Analysis
Dempster-Shafer allows results from multiple
views to be fused.
Mono-Image Belief Mono-Image Belief Mono-Image Belief Mono-Image Belief Mono-Image Belief Fused Belief Fused Belief Fused Belief Fused Belief Fused Belief
Obj Cyl Sph Cone Clutt Objs Fused Cyl Sph Cone Clutt
1 0.70 0.00 0.00 0.21 1 0.70 0.00 0.00 0.21
2 0.83 0.00 0.00 0.08 1,2 0.93 0.00 0.00 0.05
3 0.83 0.00 0.00 0.08 1,2,3 0.98 0.00 0.00 0.01
4 0.17 0.00 0.00 0.67 1,2,3,4 0.96 0.00 0.00 0.03
28
Multi-Image Analysis
Mono-Image Belief Mono-Image Belief Mono-Image Belief Mono-Image Belief Mono-Image Belief Fused Belief Fused Belief Fused Belief Fused Belief Fused Belief
Obj Cyl Sph Cone Clutt Objs Fused Cyl Sph Cone Clutt
5 0.00 0.17 0.23 0.45 5 0.00 0.17 0.23 0.45
6 0.00 0.00 0.37 0.44 5,6 0.00 0.00 0.30 0.60
7 0.00 0.303 0.45 0.045 5,6,7 0.00 0.02 0.67 0.17
8 0.00 0.32 0.23 0.31 5,6,7,8 0.00 0.01 0.62 0.20
29
Future Research
The current detection model considers objects as
a Highlight/Shadow pair. An object can also be
considered as a discrepancy in the surrounding
texture field.
30
Conclusions

• Automated Detection/Feature Extraction model has
  been developed and tested on a large amount of
  data. Good Results obtained, improvements
  expected when surface returns removed.
• Classification model uses a simple sonar
  simulator and Dempster-Shafer theory to classify
  the objects. Extends mine/not-mine
  classification to provide shape and size
  information.
• Future research is focusing on texture
  segmentation to complement the current work.

31
Acknowledgements

• We would like to thank the following institutions
  for
• their support and for providing data
• DRDCAtlantic, Canada
• Saclant Centre, Italy
• GESMA, France