Sparse reduced (spare) filter

• US 8,995,722 B2
• Filed: 08/05/2013
• Issued: 03/31/2015
• Est. Priority Date: 08/05/2013
• Status: Active Grant

• Current Assignee: Raytheon Company (Rtx Corporation)
• Original Assignee: Raytheon Company (Rtx Corporation)
• Inventors: Flanders, Bradley A., Robinson, Ian S., Sommese, Anthony M.
• Primary Examiner(s): Bitar, Nancy
• Application Number:
  US13/959,335
• Publication Number:
  US 20150036877A1
• Time in Patent Office: 603 Days
• Field of Search: None
• US Class Current: 382/103
• CPC Class Codes: G06F 17/16 Matrix or vector computation , e.g. matrix-matrix or matrix-vector multiplication, matrix factorization matrix transposition G06F7/78
• CPC Class Codes: G06F 2218/18 by plotting the signal segments against each other, e.g. analysing scattergrams
• CPC Class Codes: G06K 7/146 the method including quality enhancement steps
• CPC Class Codes: G06T 2207/10028 Range image; Depth image; 3D point clouds
• CPC Class Codes: G06T 5/20 using local operators
• CPC Class Codes: G06V 20/13 Satellite images
• CPC Class Codes: G06V 20/194 using hyperspectral data, i.e. more or other wavelengths than RGB

Abstract

The disclosure provides a filtering engine for selecting sparse filter components used to detect a material of interest (or specific target) in a hyperspectral imaging scene and applying the sparse filter to a plurality of pixels in the scene. The filtering engine transforms a spectral reference representing the material of interest to principal components space using the eigenvectors of the scene. It then ranks sparse filter components based on each transformed component of the spectral reference. The filtering engine selects sparse filter components based on their ranks. The filtering engine performs the subset selection quickly because the computations are minimized; it processes only the spectral reference vector and covariance matrix of the scene to do the subset selection rather than process a plurality of pixels in the scene, as is typically done. The spectral filter scores for the plurality of pixels are calculated efficiently using the sparse filter.

Claims

1. A method for selecting a subset of hyperspectral imaging scene spectral covariance matrix principal components to detect a material of interest or specific target in a scene, the method comprising:

in a filtering engine provided with a set of whitened principal component coefficients of a spectral reference vector, the spectral reference vector representative of a material of interest or specific target;

computing a signal-to-clutter ratio (SCR) of the spectral reference vector based on the set of whitened principal component coefficients of the spectral reference vector;

ranking the contribution of each whitened principal component coefficient to the total SCR;

selecting, based on the ranking, a subset of whitened principal component coefficients from the set of whitened principal component coefficients of the spectral reference vector;

for each of a plurality of scene pixels in the scene, computing a sparse detection filter score for a subject scene pixel based on the selected subset of whitened principal component coefficients of the spectral reference vector and a respective subset of whitened principal component coefficients of the subject scene pixel having the same indices; and

determining, based on the sparse detection filter score, whether the material of interest or specific target is present in the subject scene pixel.

2. The method of claim 1 wherein selecting the subset of whitened principal component coefficients includes selecting a set of whitened principal component coefficients of the spectral reference vector having a cumulative SCR contribution above a threshold.

3. The method of claim 1 wherein selecting the subset of whitened principal component coefficients includes selecting a set of whitened principal component coefficients of the spectral reference vector having a SCR contribution above a threshold percentage of the total SCR.

4. The method of claim 1 wherein computing the sparse detection filter score for the subject scene pixel includes computing the dot product of the selected subset of whitened principal component coefficients of the spectral reference vector with the respective subset of whitened principal component coefficients of the subject scene pixel having the same indices.

5. The method of claim 1 wherein determining whether the material of interest or specific target is present in the subject scene pixel includes:

comparing the sparse detection filter score for the subject scene pixel to a threshold; and

determining the presence of the material of interest or specific target in the subject scene pixel based on the comparison.

6. The method of claim 5 wherein scene pixels determined to have the material of interest or specific target present are initial detections, the method further comprising:

computing full dimension matched filter (MF), Adaptive Cosine/Coherence Estimator (ACE) or other spectral detection filter scores for verifying the initial detections; and

determining, based on the full dimension matched filter (MF), Adaptive Cosine/Coherence Estimator (ACE) or other spectral detection filter scores, in which of the initial detections is the material of interest or specific target present.

7. The method of claim 1 further comprising:

comparing the sparse detection filter scores to a threshold;

selecting scene pixels from the plurality of scene pixels based on the comparison;

computing a modified detection filter score for each of the selected scene pixels based on a local spectral covariance in the vicinity of the material of interest or specific target; and

determining, based on the modified detection filter scores, in which of the selected scene pixels is the material of interest or specific target present.

8. The method of claim 1, given M whitened principal component coefficients selected for the spectral reference vector, and wherein the subject scene pixel includes a plurality of principal components and all the principal components of the plurality have been unmixed from the subject scene pixel and whitened, the method further comprising:

from the plurality, selecting whitened principal component coefficients of the subject scene pixel having the same indices as the whitened principal component coefficients of the spectral reference vector based on the M whitened principal component coefficients selected for the spectral reference vector.

9. The method of claim 1, given M whitened principal component coefficients selected for the spectral reference vector and M principal components from which the coefficients are computed, and given a set of zero or more previously unmixed whitened principal component coefficients for the subject scene pixel, the method further comprising:

from the set, selecting previously unmixed whitened principal component coefficients having the same indices as the whitened principal component coefficients selected for the spectral reference vector, based on the M whitened principal component coefficients;

for each coefficient of the M whitened principal component coefficients having an index different than the indices of the previously unmixed whitened principal component coefficients, unmixing a principal component, which is associated with that coefficient, from the subject scene pixel resulting in a newly unmixed whitened principal component coefficient for the subject scene pixel; and

wherein computing the sparse detection filter score includes computing the sparse detection filter score for the subject scene pixel using the previously and newly unmixed whitened principal component coefficients having the same indices as the M whitened principal component coefficients selected for the spectral reference vector.

10. The method of claim 9 further comprising for the subject scene pixel, recording the newly unmixed whitened principal components coefficients with the previously unmixed whitened principal components coefficients.

11. The method of claim 9, given M′

whitened principal component coefficients selected for a second spectral reference vector and M′

principal components from which the coefficients are computed, the method further comprising;

from the set, selecting previously unmixed whitened principal component coefficients having the same indices as the whitened principal component coefficients selected for the second spectral reference vector, based on the M′

whitened principal component coefficients;

for each coefficient of the ‘

M whitened principal component coefficients having an index different than the indices of the previously unmixed whitened principal component coefficients, unmixing a principal component, which is associated with that coefficient, from the subject scene pixel resulting in a newly unmixed whitened principal component coefficient for the subject scene pixel; and

with respect to the second spectral reference vector, computing a sparse detection filter score for the subject scene pixel using the previously and newly unmixed whitened principal component coefficients having the same indices as the M′

whitened principal component coefficients selected for the second spectral reference vector.

12. A system for selecting a subset of hyperspectral imaging scene spectral covariance matrix principal components to detect a material of interest or specific target in a scene, the system comprising:

a memory having computer executable instructions thereupon;

at least one interface receiving a set of whitened principal component coefficients of a spectral reference vector, the spectral reference vector representative of a material of interest or specific target;

a filtering engine coupled to the memory and the at least one interface, the computer executable instructions when executed by the filtering engine cause the filtering engine to;

compute a signal-to-clutter ratio (SCR) of the spectral reference vector based on the set of whitened principal component coefficients of the spectral reference vector;

rank the contribution of each whitened principal component coefficient to the total SCR;

select, based on rank, a subset of whitened principal component coefficients from the set of whitened principal component coefficients of the spectral reference vector; and

for each of a plurality of scene pixels in the scene, compute a sparse detection filter score for a subject scene pixel based on the selected subset of whitened principal component coefficients of the spectral reference vector and a respective subset of whitened principal component coefficients of the subject scene pixel having the same indices; and

determine, based on the sparse detection filter score, whether the material of interest or specific target is present in the subject scene pixel.

13. A tangible non-transitory computer-readable storage medium having computer readable instructions stored therein for selecting a subset of hyperspectral imaging scene spectral covariance matrix principal components to detect a material of interest or specific target in a scene, which when executed by one or more processors provided with a set of whitened principal component coefficients of a spectral reference vector, the spectral reference vector representative of a material of interest or specific target:

compute a signal-to-clutter ratio (SCR) of the spectral reference vector based on the set of whitened principal component coefficients of the spectral reference vector;

rank the contribution of each whitened principal component coefficient to the total SCR;

select, based on rank, a subset of whitened principal component coefficients from the set of whitened principal component coefficients of the spectral reference vector; and

for each of a plurality of scene pixels in the scene, compute a sparse detection filter score for a subject scene pixel based on the selected subset of whitened principal component coefficients of the spectral reference vector and a respective subset of whitened principal component coefficients of the subject scene pixel having the same indices; and

determine, based on the sparse detection filter score, whether the material of interest or specific target is present in the subject scene pixel.