(Automatic) Target Detection in Synthetic Aperture Radar Imagery Via Terrain Recognition

1
(Automatic) Target Detection in Synthetic
Aperture Radar Imagery Via Terrain Recognition

• Authors Rupert Paget, John Homer, and David Crisp

THE UNIVERSITY OF QUEENSLAND AUSTRALIA
Cooperative Research Centre for Sensor Signal
and Information Processing
2
Contents

• The problem
• Markov random field texture model
• Open ended texture classification
• Target detection
• The results
• Conclusion

3
The Problem

• To identify real targets from background texture.
• Surveillance of large areas of the earths surface
  is often undertaken with low resolution synthetic
  aperture radar (SAR) imagery from either a
  satellite or a plane.
• There is a need to process these images with
  automatic target detection (ATD) algorithms.

Identified real targets
False targets
4
The Problem

• Typically the targets being searched for are
  vehicles or small vessels, which occupy only a
  few resolution cells.
• Simple thresholding is usually inadequate for
  detection due to the high amount of noise in the
  images.
• Often the background has a discernible texture,
  and one form of detection is to search for
  anomalies in the texture caused by the presence
  of the target pixels.

Identified real targets
False targets
5
The Problem

• To perform this task a texture model must be able
  to model a variety of textures at run time, and
  also model these textures well enough to detect
  anomalies.
• We accomplish this with our multiscale
  nonparametric Markov random field (MRF) texture
  model.

Identified real targets
False targets
6
Markov Random Field Model

• Is formed by modelling the value of the centre
  pixel in terms of a conditional probability with
  respect to its neighbouring pixels values.

7
Nonparametric MRF Model

• Built from a multidimensional histogram.
• Does not require parameter estimation.
• Can model high dimensional statistics.

8
Strong Nonparametric MRF

• Where the multidimensional histogram is
  represented as a combination of marginal
  histograms.
• This allows control over the statistical order of
  the model.

9
Synthetic Textures

• Comparative analysis of the synthetic textures
  shows that the texture model can capture the
  unique characteristics of various textures.

10
Open Ended Classification

• To perform target detection, or anomaly
  detection, we will use our open ended texture
  classifier.
• It is based on the notion that if a texture model
  is able to capture the unique characteristics of
  a texture, then the distribution of those
  characteristics or features define the texture.

Conventional N class classifier
Open ended classifier
11
Open Ended Classification

• A texture is classified if it has the same set of
  characteristics or features as a predefined
  texture.
• This is resolved via a goodness-of-fit test
  between the two sets of characteristics.
• Such a method allows the unknown or uncommitted
  subspace to be left undefined.

Conventional N class classifier
Open ended classifier
12
Goodness-of-fit Test

• Require a population of measurements.
• Most reliable results are from one-dimensional
  statistics.
• Therefore
• We use the nonparametric model to obtain
  histograms, using the data points as features or
  measurements. This gives us a population of
  measurements.
• To obtain one-dimensional statistics from a
  multi-dimensional histogram, we discard the
  positional information and just use the
  frequencies or probabilities or distance to the
  nearest neighbour associated with the data points.

13
Target Detection

• Given that the images have been pre-segmented, we
  wish to determine whether there is a target in
  the centre of some undefined texture.
• First, build the histograms for the nonparametric
  MRF model of the background texture.
• For each histogram, create a set of one
  dimensional statistics for both background
  texture and target.
• These sets of one dimensional statistics can
  again be reduced to just one set of one
  dimensional statistics.
• Perform a goodness-of-fit on this set of
  statistics. We used the nonparametric
  Kruskal-Wallis test.

14
Results
MRF Model True Targets False Targets Difference
n1c0t0w2 88.5167 12.5846 75.9321
n1c0t0w4 94.0191 12.5056 81.5135
n1c0t0w6 93.5407 11.7728 81.7679
n1c0t1w2 60.5263 33.1926 27.3337
n1c0t1w4 82.2967 39.2159 43.0808
n1c0t1w6 86.6029 38.6314 47.9715
n1c2t0w2 93.5407 16.7668 76.7739
n1c2t0w4 97.6077 24.9306 72.6771
n1c2t0w6 95.6938 21.5264 74.1674
n1c2t1w2 31.1005 22.0496 9.05090
n1c2t1w4 87.0813 43.5676 43.5137
n1c2t1w6 84.2105 29.9600 54.2505

• Nearest neighbour neighbourhood nonparametric MRF
  models with their best target discrimination
  performance.

15
Results
MRF Model True Targets False Targets Difference
n3c0t0w2 84.6890 10.6531 74.0359
n3c0t0w4 96.6507 18.9497 77.7010
n3c0t0w6 93.7799 14.7908 78.9891
n3c0t1w2 54.0670 27.4947 26.5723
n3c0t1w4 83.7321 38.4853 45.2468
n3c0t1w6 84.4498 33.2018 51.2480
n3c2t0w2 95.6938 26.4195 69.2743
n3c2t0w4 99.7608 46.1267 53.6341
n3c2t0w6 97.8469 35.9212 61.9257
n3c2t1w2 60.7656 40.0080 20.7576
n3c2t1w4 80.1435 23.8357 56.3078
n3c2t1w6 85.4067 24.3666 61.0401

• 3x3 neighbourhood nonparametric MRF models with
  their best target discrimination performance.

16
Results
MRF Model True Targets False Targets Difference
n0t0w2 79.6651 13.0061 66.6590
n0t0w4 87.7990 14.1505 73.6485
n0t0w6 84.4498 9.27080 75.1790
n0t1w2 46.4115 30.5805 15.8310
n0t1w4 51.1962 21.4410 29.7552
n0t1w6 83.7321 33.3387 50.3934
Histograms True Targets False Targets Difference
t0w2 80.6220 16.7287 63.8933
t0w4 94.0191 40.9498 53.0693
t0w6 86.1244 37.3967 48.7277
t1w2 99.0431 54.9217 44.1214
t1w4 98.0861 51.3008 46.7853
t1w6 84.9282 37.8322 47.0960

• Control models with their best target
  discrimination performance.

17
Conclusion

• The results were obtained from a DSTO data set
  containing 142067 pre-segmentated images of
  possible targets. 418 of these images were ground
  truthed as having real targets.
• Our best results were able to reduce the number
  of false targets to 11.8 while retaining 93.5
  of the true targets.
• This texture discrimination method was shown to
  be better than comparable grey level
  discrimination.

18
Conclusion

• Future direction of this research is to increase
  the speed of the algorithm. This may require new
  discriminating features.
• This will allow implementation of the algorithm
  on a larger DSTO target detection database.
• From these future results we will be able to
  compare our method with current target detection
  methods.