The Automatic Target Recognition System in SAIP

Authored by: Leslie M. Novak, Gregory J. Orwika, William S. Brower, and Alison L. Wever

Published by:

The Lincoln Laboratory Journal, 1997

Presented by: Umur Kathree

Automatic Target Recognition

Goal of Automatic Target Recognition (ATR) is to detect and recognise targets in images produced by SAR.

Current ATR algorithms are not fully automatic i.e there would always be a need for human operators.

Advances in surveillance and reconnaissance technologies would increase the demands on image analysts.

CSIR 2012

The SAIP System

The SAIP System

The Semi-Automated IMINT Processing (SAIP) project has been developed by DARPA to address the growing quantities of image data. The system would require fewer image analysts to examine the data in near real time.

It would provide visualisation cues to help in the interpretation of the images.

Uses a superresolution technique known as High-Definition Vector Imaging (HDVI).

CSIR 2012

Figure 1: Process flow for phase history data gathered by an unmanned air vehicle

The SAIP System

The SAIP consists of a CFAR detector that locates areas in the SAR image having high radar contrast.

These areas, called chips, are then extracted from the larger SAR image.
The chips could be the targets of interest or false alarms from the clutter environment. They are then sent to the feature extractor where sets of features such as length, width and diameter are calculated.

CSIR 2012

The SAIP System

The feature extractor produces a scored measure of target likeness from the calculated features.

There are other false-alarm mitigation modules, such as terrain delimitation, that also provide target likeness scores.

All these different scores are weighted and combined into an overall score. They are sent into the discrimination-thresholding module where candidate targets are thresholded and prioritized for further processing in the HDVI-MSE classifier

CSIR 2012

The SAIP System

The HDVI processing is applied to the images of the candidate targets to enhance their resolutions before classification. They are then compared to a set of templates where the mean squared error (MSE) between them is calculated. The minimum MSE score over aspect angle is converted to a confidence value for each target type. These confidence values, as well as the HDVI processed images and corresponding templates, are reviewed by the image analysts for target recognition.

CSIR 2012

Figure 2: Process flow for SAIP system

Overview of paper
To quantify the improvement in target recognition performance using HDVI before classification. To introduce a two-stage HDVI-MSE classifier that would improve the computational speed at a cost of some decrease in performance. To evaluate the end-to-end ATR performance using the SAIP detection, discrimination and HDVI classifier stages.

CSIR 2012

Description of Data

Description of Data
The SAR images were provided by Wright Laboratories. These data were gathered at the Redstone Arsenal in Huntsville, Alabama using a Sandia X-band SAR sensor. The data comprises of military targets imaged over 360 of aspect angle in spotlight mode, and 30 km2 of clutter in stripmap mode. Ten targets were used to train the classifier by constructing templates. Eight targets were used to test it. The HMMWV and the M35 were used as confuser vehicles and should be classified as unknown.

CSIR 2012

Figure 3: The eighteen vehicle set used to test the SAIP. The training targets were the BMP2 #1, M2 #1, T72 #1, BTR60, BTR70, M1, M109, M110, M113 and the M548. The rest were used to train the SAIP. The HMMWV and the M35 were the confuser vehicles.

Description of Data
The data would be presented in terms of classifier confusion matrices (tables) that show the number of correct and incorrect classifications. They are then summarised in terms of a probability of correct classification (Pcc) based on the total number of targets entered into the system. This probability includes the number of confuser vehicles correctly identified as unknown.

CSIR 2012

Performance results for NonHDVI-Processed Data

Non-HDVI-Processed Data
To test the classifier with conventional SAR images to establish a baseline for testing performance. Three nominal resolutions were used: 0.3m x 0.3m, 0.5m x 0.5m and 1m x 1m.

At each resolution of the target data, the MSE classifier was implemented and tested using 72 templates per target.

CSIR 2012

The Pcc was found to be at 93.9%.

The Pcc was found to be at 84.1%.

The Pcc was found to be at 45.4%.

Non-HDVI-Processed Data

Based on the previous results, a resolution of 0.5m x 0.5m or better is needed to achieve a reliable probability of correct classification.

CSIR 2012

Performance results for HDVIProcessed Data

HDVI-Processed Data
The HDVI-MSE classifier was implemented to compare the performance of HDVI-processed data to conventional SAR data. Two nominal resolutions of 0.3m x 0.3m and 1.0m x 1.0m were used. After HDVI processing, they were reprocessed into resolutions of 0.15m x 0.15m and 0.5m x 0.5m respectively.

CSIR 2012

Figure 4: SAR images of an M35 truck. The left image has been processed with conventional techniques and has a resolution of 0.3m x 0.3m, while the right image has undergone HDVI processing resulting in a resolution of 0.15m x 0.15m.

The Pcc was found to be at 96.4%.

The Pcc was found to be at 73.4%.

HDVI-Processed Data

HDVI processing allows for a reliable Pcc for SAR images as coarse as 1.0m x 1.0m resolution.

CSIR 2012

The Two-Stage Classifier

Two-stage classifier
The two-stage classifier is implemented with a preceding preclassifier stage that performs a coarse MSE classification on non-HDVI data. The preclassifier would provide an estimate of the targets aspect angle as well as its class. This information is then passed to the HDVI-MSE classifier and used to limit its template search space resulting in a computationally efficient algorithm.

CSIR 2012

Figure 5: Comparison of the two classifier schemes.

The Pcc was found to be at 70.0%.

Two-stage classifier
To begin estimating the computational speed of this scheme, it was found that the HDVI-MSE classifier required 81 times more computations than an MSE classifier on conventional SAR images. Then it was found that for the HDVI-MSE classifier with a reduced search space, the number of computations required were reduced by a factor of 50. Therefore, in total, the two-stage classifier was found to be around 30 times faster than the single-stage HDVI-MSE classifier.

CSIR 2012

End-to-End Performance

End-to-End Performance
The system implementation consisted of a two parameter CFAR detector, a discrimination stage and the two-stage classifier. The data sets used were the independent test target images and the 30 km2 clutter set. The resolution used was 1.0m x 1.0m. Performance results are presented in terms of ROC curves showing probability of detection versus false-alarm density (false alarms/km2)

CSIR 2012

End-to-End Performance
As was mentioned previously, the CFAR detector extracts localised regions in the SAR image displaying high radar contrast. These could either be targets of interest or clutter false alarms. As will be shown in the next slide, selecting a CFAR detection threshold of 3,5 would result in a 0.95 detection probability and around 30 false alarms/ km2. This is around 905 false objects detected in the 30 km2 area.

CSIR 2012

Figure 6: ROC performance curves for each element of the end-to-end ATR system

End-to-End Performance
The discriminator module uses two measures of target likeness to separate targets of interest from clutter false alarms. Firstly, a set of features are calculated and compare to the training data to provide a quadratic distance (QD) score. Then, inner products of the vector of features and a weight vector are computed to provide a quadratic polynomial discriminator (QPD) score.

CSIR 2012

Figure 7: Distribution of QD scores. Selecting a threshold of 13.1 would provide a detection probability of 0.95 at 21.4 false alarms/km2 (263 out of 905 rejected false alarms)

Figure 8: Distribution of QPD scores. Selecting a threshold of 0.26 would provide a 0.95 detection probability and a reduced 12 false alarms/km2 (537 out of 905 rejected false alarms)

End-to-End Performance
Referring back to Figure 6, the QD and QPD scores were combined and are presented by the blue curve. Selecting the operating point at 0.9 detection probability reduced the false alarm density to 1 false alarm/km2 (875 out of 905 rejected false alarms) In testing out the two-stage classifier, at a detection probability of 0.8, all but 3 clutter false alarms were rejected (0.1 false alarms/km2) This would also be shown by the red curve in Figure 6

CSIR 2012

Summary
A resolution of 0.5m x 0.5m or better was needed for reliable Pcc. But with HDVI processing, this can be achieved even with a 1.0m x 1.0m resolution image. The two-stage classifier was 30 times faster than the single stage HDVI-MSE classifier, but at a cost of marginal decrease in ATR performance. For end-to-end performance, 0.1 false alarms/km2 could be achieved at a detection probability of 80%.

CSIR 2012

References
1. W.P. Delaney, The Changing World, the Changing Nature of Conflicts: A Critical Role for Military Radar, 1995 IEEE Natl. Radar Conf., Alexandria, Va., 811 May 1995, pp. 1115. 2. D.A. Fulghum, DARPA Looks Anew at Hidden Targets, Aviation Week Space Technol., 6 Jan. 1997, pp. 5657. 3. G.R. Benitz, High-Definition Vector Imaging for Synthetic Aperture Radar, Proc. 31st Asilomar Conf. on Systems, Signals & Computers 2, Pacific Grove, Calif., 25 Nov. 1997, pp. 12041209. 4. L.M. Novak, S.D. Halversen, G.J. Owirka, and M. Hiett, Effects of Polarization and Resolution on the Performance of a SAR ATR System, Linc. Lab. J. 8 (1), 1995, pp. 4968. 5. L.M. Novak, G.R. Benitz, G.J. Owirka, and L.A. Bessette, ATR Performance Using Enhanced Resolution SAR, SPIE 2757, 1996, pp. 332337.

6. E.R. Keydel and S.W. Lee, MSTAR Extended Operating Conditions: A Tutorial, SPIE 2757, 1996, pp. 228242.
7. G.J. Owirka, A.L. Weaver, and L.M. Novak, Performance of a Multi-Resolution Classifier Using Enhanced Resolution SAR Data, SPIE 3066, 1997, pp. 90100. 8. J.C. Principe, A. Radisavljevic, M. Kim, J. Fisher III, M. Hiett, and L.M. Novak, Target Prescreening Based on 2D Gamma Kernels, SPIE 2487, 1995, pp. 251258.

CSIR 2012