Fusion of shape and multiscale features for unknown target rejection

• US 6,897,802 B1
• Filed: 11/10/2003
• Issued: 05/24/2005
• Est. Priority Date: 11/10/2003
• Status: Expired due to Term

• Current Assignee: Raytheon Company (Rtx Corporation)
• Original Assignee: Raytheon Company (Rtx Corporation)
• Inventors: Daniell, Cynthia, Srinivasa, Narayan
• Primary Examiner(s): Sotomayor, John B.
• Application Number:
  US10/704,841
• Time in Patent Office: 561 Days
• Field of Search: 342/90, 342/93, 342/91, 342/89, 342/25.A, 342/25.F, 342/195, 342/191, 342/190, 342/179, 342/176, 342/175
• US Class Current: 342/90
• CPC Class Codes: G01S 13/9005 with optical processing of the SAR signals
• CPC Class Codes: G01S 13/9027 Pattern recognition for feature extraction

Abstract

A plurality of image chips (202) (over 100), each of the chips containing the same, known target of interest, such as, for example an M109 tank are presented to the system for training. Each image chip of the known target is slightly different than the next, showing the known target at different aspect angles and rotation with respect to the moving platform acquiring the image chip.

The system extract multiple features of the known target from the plurality of image chips (202) presented for storage and analysis, or training. These features distinguish a known target of interest from the nearest similar target to the M109 tank, for example a Caterpillar D7 bulldozer. These features are stored for use during unknown target identification. When an unknown target chip is presented, the recognition algorithm relies on the features stored during training to attempt to identify the target.

The tools used for extracting features of the known target of interest as well as the unknown target presented for identification are the same and include the Haar Transform (404), and entropy measurements (410) generating coefficient locations. Using the Karhunen-Loeve (KL) transform 406, eigenvectors are computed. A Gaussian mixture model (GMM) (507) is used to compare the extracted coefficients and eigenfeatures from the known target chips with that of the unknown target chips. Thus the system is trained initially by presenting to it known target chips for classification. Subsequently, the system uses the training in the form of stored eigenfeatures and entropy coefficients fused with multiscale features to identify unknown targets.

Claims

1. A method for automatic target recognition, said target acquired as part of a radar image, said radar image formed from digitized returns, said digitized returns processed into pixels forming said radar image, each of said pixels having an amplitude, comprising the steps of:

storing said pixels forming said radar image in a memory,

pre-processing said pixels forming said radar image to extract a target chip containing said target from said image;

applying a first recognition algorithm to said target chip to identify a first classification of said target extracted from said image;

applying a second recognition algorithm to said target chip to identify a second classification of said target extracted from said image, said second algorithm complementary to said first algorithm;

fusing said first classification and said second classification to generate a target classification identifying said target.

2. A method as described in claim 1 wherein said pre-processing of said image also extracts a shadow cast by said target.

3. A method as described in claim 2 wherein said first algorithm performs a multiscale analysis.

4. A method as described in claim 3 wherein said multiscale analysis is not dependent on segmentation.

5. A method as described in claim 4 wherein said multiscale analysis is not invariant to target scale, target translation and target rotation.

6. A method as described in claim 2 wherein said second algorithm performs a shape statistics analysis.

7. A method as described in claim 6 wherein said said shape statistics analysis is invariant to scale.

8. A method as described in claim 6 wherein said shape statistics analysis is invariant to target translation, rotation and scaling.

9. A method as described in claim 3 wherein said pre-processing comprises the steps of:

computing a mean and a standard deviation of said target chip containing the target using the amplitudes of each pixel part of said target chip;

computing a first threshold equal to said two standard deviations below said mean of said target chip;

setting to zero all pixels within said target chip having said amplitude below said first threshold to create a first thresholded target chip;

computing a second mean and a second standard deviation using amplitude of pixels part of said first thresholded target chip;

computing a second threshold, said second threshold one-half of said second standard deviation above said second mean;

setting to zero all pixels within said first thresholded target chip below said second threshold to generate a second thresholded chip;

subtracting an amplitude of the minimum non-zero value from each column of said second thresholded chip to generate a third thresholded target chip;

performing a binarization of said third thresholded target chip using morphological filtering of holes and single pixel noise to generate a fourth target chip;

performing a Haar transform on said fourth target chip to generate a transformed target chip.

10. A method as described in claim 9 wherein said Haar transform is performed for three levels.

11. A method as described in claim 10 wherein said first algorithm tests said target chip comprising the steps of:

extracting the entropy features of said target from stored coefficient locations;

computing one or more eigenfeatures said eigenfeatures uniquely distinguishing said target;

appending the entropy features and eigenfeatures;

computing distances of said eigenfeatures extracted from said target chip;

thresholding said distances to limits identifying said target.

12. A method as described in claim 11 wherein said second algorithm is adaptive to a new target by accepting a new set of eigenfeatures.

13. An apparatus for automatic target recognition, said target acquired as part of a radar image, said radar image formed from digitized returns, said digitized returns processed into pixels forming said radar image, each of said pixels having an amplitude, comprising:

memory for storage of said pixels forming said radar image;

a processor for pre-processing said pixels forming said radar image to extract a target chip containing said target from said image;

said processor applying a first recognition algorithm to said target chip to identify a first classification of said target extracted from said image;

said processor applying a second recognition algorithm to said target chip to identify a second classification of said target extracted from said image, said second algorithm complementary to said first algorithm;

said processor fusing said first classification and said second classification to generate a target classification identifying said target.

14. An apparatus as described in claim 13 wherein said processor also extracts a shadow cast by said target from said radar image.

15. An apparatus as described in claim 14 wherein said first recognition algorithm performs a multiscale analysis.

16. A method as described in claim 15 wherein said multiscale analysis is not dependent on segmentation.

17. An apparatus as described in claim 16 wherein said multiscale analysis is not invariant to target scale, target translation and target rotation.

18. An apparatus as described in claim 14 wherein said second recognition algorithm performs a shape statistics analysis.

19. An apparatus as described in claim 18 wherein said said shape statistics analysis is invariant to scale.

20. An apparatus as described in claim 19 wherein said shape statistics analysis is invariant to target translation, rotation and scaling.

21. An apparatus as described in claim 15 wherein said processor performs the steps of:

computing a mean and a standard deviation of said target chip containing the target using the amplitudes of each pixel part of said target chip;

computing a first threshold equal to said two standard deviations below said mean of said target chip;

setting to zero all pixels within said target chip having said amplitude below said first threshold to create a first thresholded target chip;

computing a second mean and a second standard deviation using amplitude of pixels part of said first thresholded target chip;

computing a second threshold, said second threshold one-half of said second standard deviation above said second mean;

setting to zero all pixels within said first thresholded target chip below said second threshold to generate a second thresholded chip;

subtracting an amplitude of the minimum non-zero value from each column of said second thresholded chip to generate a third thresholded target chip;

performing a binarization of said third thresholded target chip using morphological filtering of holes and single pixel noise to generate a fourth target chip;

performing a Haar transform on said fourth target chip to generate a transformed target chip.

22. An apparatus as described in claim 21 wherein said Haar transform is performed for three levels.

23. An apparatus as described in claim 21 wherein said first algorithm performed by said processor comprises the steps of:

extracting the entropy features of said target from stored coefficient locations;

computing one or more eigenfeatures said eigenfeatures uniquely distinguishing said target;

appending the entropy features and eigenfeatures;

computing distances of said eigenfeatures extracted from said target chip;

thresholding said distances to limits identifying said target.

24. A method as described in claim 21 wherein said second algorithm is adaptive to a new target by accepting a new set of eigenfeatures.