System and method for adaptive face recognition

• US 8,885,893 B1
• Filed: 11/16/2012
• Issued: 11/11/2014
• Est. Priority Date: 11/22/2011
• Status: Active Grant

• Current Assignee: the united states of america as represented by the secretary of the navy
• Original Assignee: the united states of america as represented by the secretary of the navy
• Inventors: Estabridis, Katia N.
• Primary Examiner(s): Park, Chan
• Assistant Examiner(s): Roz, Mark
• Application Number:
  US13/679,488
• Time in Patent Office: 725 Days
• Field of Search: None
• US Class Current: 382/118
• CPC Class Codes: G06V 10/462 Salient features, e.g. scale invariant feature transforms [SIFT]
• CPC Class Codes: G06V 40/172 Classification, e.g. identification

Abstract

A system and method for adaptive face recognition includes at least one electronic processor having a central processing unit. At least one database having a plurality of pixilated face images of known subjects of interest is associated with the processor. At least one test image of a new subject of interest is configured for input into the electronic processor. A classification processing tool is associated with the electronic processor. The classification processing tool is configured to build a dictionary and provide a classification match of the test image with one of the plurality of pixilated face images of known subjects of interest. At least one device is associated with the processor and configured to output the classification match in a tangible medium.

Claims

1. An adaptive face recognition system, comprising:

at least one electronic processor having a central processing unit;

at least one database having a plurality of pixilated face images of known subjects of interest, wherein said at least one database is associated with said at least one electronic processor;

at least one test image, j, of a new subject of interest, wherein said at least one test image is a pixilated face image configured for input into said at least one electronic processor;

a classification processing tool associated with said at least one electronic processor, wherein said classification processing tool is configured to build a dictionary, A, associated with said at least one database, wherein said classification processing tool is configured to provide a classification match of said at least one test image, j, with one of said plurality of pixilated face images of known subjects of interest; and

at least one device associated with said at least one electronic processor configured to output in a tangible medium said classification match;

wherein said classification processing tool is a non-transitory electronic-processor-readable medium having a plurality of electronic processor executable instructions stored thereon, that when executed by said at least one electronic processor, causes said at least one electronic processor to;

input said at least one test image, j, into said at least one electronic processor;

perform adaptive high-pass filtering, HF, on said at least one test image, j, to increase the number of keypoints, KP, in said at least one test image, j, to reduce changes in illumination in said at least one test image, j, wherein said keypoints, KP, are structures of interest in said at least one test image, j, including eyes, brows, nose, mouth, and hairline features, wherein said adaptive high-pass filtering, HF, uses at least two high-pass filters having different spatial frequency bandwidths dependent on the pixel size of said at least one test image, j;

compute at least one scale invariant feature transform (SIFT) descriptor, N, for each of said at least one test image, j;

recognize said new subject of interest in said at least one test image, j, by using an optimization technique, l₁, to search for a best match of said new subject of interest in said at least one test image, j, in said at least one database having a plurality of pixilated face images of known subjects of interest;

determine whether to learn and alter the current knowledge of said dictionary, A from A_c(j-1) to A_c(j);

output a class having maximum output, q_max;

when it is determined to learn, adapt said dictionary, A, using reversed-orthogonal matching pursuit (OMP); and

iterate through said recognize said new subject of interest in said at least one test image, j, by using an optimization technique, l₁, to search for a best match of said new subject of interest in said at least one test image, j, in said at least one database having a plurality of pixilated face images of known subjects of interest task, said determine whether to learn and alter the current knowledge of said dictionary, A task, and said output a class having maximum output, q_max, until it is determined not to learn.

2. The system according to claim 1, wherein said adaptive high-pass filtering task, comprises:

passing said at least one test image, j, through said at least two adaptive high-pass filters, HF;

extracting a plurality of keypoints, KP, from said at least one test image, j, by filtering said at least one test image, j, with each of said at least two adaptive high-pass filters, HF, wherein said extracting is performed by lowering the threshold of strength, t, wherein t=t−

c, wherein c is a user-selected constant based on application-specific conditions;

selecting, in each pass through each of said at least two adaptive high-pass filters, HF, at least one keypoint, KP, in said at least one test image, j;

determining, in each pass through each of said at least two adaptive high-pass filters, HF, whether the number of keypoints, KP, is greater than or equal to the number of said descriptors, N, wherein said determination is expressed as KP≧

N;

when, in each pass through each of said at least two adaptive high-pass filters, said number of keypoints, KP, is not greater than or equal to the number of said at least one SIFT descriptor, N, iterating through said extracting a plurality of keypoints, KP, from said at least one test image, j, wherein said extracting is performed by lowering the threshold of strength, t, wherein t=t−

c, wherein c is a user-selected constant based on application-specific conditions, until said number of keypoints KP is greater than or equal to the number of said at least one SIFT descriptor, N;

extracting, in each pass through each of said at least two adaptive high-pass filters, HF, said at least one SIFT descriptor, N, from said at least one test image, j; and

summing said at least one SIFT descriptor, N, from each pass through each of said at least two adaptive high-pass filters, HF, to yield a set of N-SIFT descriptors.

3. The system according to claim 2, wherein said recognition task, comprises:

solving a matrix, X, by using a sparse solution, wherein said matrix X is said sparse solution for each of said at least one SIFT descriptor N in columns of matrix Y, wherein Y=AX; and

summing number of matches per class, wherein said number of matches per class is vector q, wherein said vector q is the vector solution extracted from said matrix X for each class, c.

4. The system according to claim 3, wherein said determination whether to learn task, comprises:

initializing model parameters for said summed number of matches per class, q;

initializing the number of clusters to a user defined number, K, wherein each of said cluster is a Gaussian distribution selected from the group of matches consisting of said summed number of matches per class vector, q;

initializing the number of iterations, M, to a user selected number to update said model parameters and each of said Gaussian distribution, and initializing the actual number of iterations performed, N_iter, to 1 on the first iteration of said number of iterations, M, wherein N_iter increases by 1 each time through said iteration;

updating model parameters for each potential k cluster,

determining whether the actual number of iterations performed, N_iter, is greater than M;

when the number of actual iterations performed, N_iter, is not greater than M, iterating through said updating model parameters for each potential k cluster task, and said determining whether the actual number of iterations performed, N_iter, is greater than M, until the actual number of iterations performed, N_iter, is greater than M;

finding a cluster with the highest mean value, μ

_c;

when said cluster with the highest mean value, μ

_c, does not belong in the same class as q_max, learning does not occur and said dictionary A_c(j-1) is not updated;

when said cluster with the highest mean value, μ

_c, belongs in the same class as q_max, determine whether said cluster with the highest mean value, μ

_c, has a single member,

wherein when said cluster with the highest mean value, μ

_c, does not have a single member, learning does not occur and said dictionary A is not updated; and

wherein when said cluster with the highest mean value, μ

_c, does have a single member, learning does occur and said dictionary A_c(j-1) is updated to A_cj, with the columns of matrix Y.

5. The system according to claim 4, wherein said adapting said dictionary, A, using reversed-orthogonal matching pursuit (OMP), comprises:

inputting said columns of updated matrix, Y_j, for said dictionary, A_c(j-1), and q_max into a matrix X_j by using a sparse solution, wherein said matrix X_j is said sparse solution, wherein A_c(j-1)=Y_jX_j, wherein Y_j is the set of N-SIFT descriptors extracted from said at least one test image, j;

finding a complement of X_j, wherein Y_j=Y_j(X_j)∪

Y_j(X_j^c), wherein X_j^c is the set of N-SIFT descriptors not used in said sparse solution; and

updating said dictionary A_c(j-1) to A_c(j), wherein A_cj=[A_c(j-1) Y(X_j^c)].

6. The system according to claim 5, wherein said outputting a class having the maximum output, q_max, comprises:

normalizing said number of matches per class, vector q, to find said class with the maximum energy, q_max; and

outputting said class having maximum energy, q_max.

7. The system according to claim 1, wherein said tangible medium is a visual display screen.

8. A method for adaptive face recognition using an electronic processor, comprising:

providing at least one database having a plurality of pixilated face images of known subjects of interest;

providing at least one test image, j, of a new subject of interest, wherein said at least one test image is a pixilated face image configured for input into at least one electronic processor,

providing a classification processing tool, wherein said classification processing tool is configured to build a dictionary, A, associated with said at least one database, wherein said classification processing tool is configured to provide a classification match of said at least one test image, j, with one of said plurality of pixilated face images of known subjects of interest; and

outputting in a tangible medium said classification match;

wherein said classification processing tool is a non-transitory electronic-processor-readable medium having a plurality of electronic processor executable instructions stored thereon, that when executed by an electronic processor, causes said at least one electronic processor to;

input said at least one test image, j, into said at least one electronic processor;

perform adaptive high-pass filtering, HF, on said at least one test image, j, to increase the number of keypoints, KP, in said at least one test image, j, to reduce changes in illumination in said at least one test image, j, wherein said keypoints, KP, are structures of interest in said at least one test image, j, including eyes, brows, nose, mouth, and hairline features, wherein said adaptive high-pass filtering, HF, uses at least two high-pass filters having different spatial frequency bandwidths dependent on the pixel size of said at least one test image, j;

compute at least one scale invariant feature transform (SIFT) descriptor, N, for each of said at least one test image, j;

recognize said new subject of interest in said at least one test image, j, by using an optimization technique, l₁, to search for a best match of said new subject of interest in said at least one test image, j, in said at least one database having a plurality of pixilated face images of known subjects of interest;

determine whether to learn and alter the current knowledge of said dictionary, A from A_c(j-1) to A_c(j);

output a class having maximum output, q_max;

when it is determined to learn, adapt said dictionary, A, using reversed-orthogonal matching pursuit (OMP); and

iterate through said recognize said new subject of interest in said at least one test image, j, by using an optimization technique, l₁, to search for a best match of said new subject of interest in said at least one test image, j, in said at least one database having a plurality of pixilated face images of known subjects of interest task, said determine whether to learn task, and said output a class having maximum output, q_max, until it is determined not to learn.

9. The method according to claim 8, wherein said adaptive high-pass filtering task, comprises:

passing said at least one test image, j, through said at least two adaptive high-pass filters, HF;

extracting a plurality of keypoints, KP, from said at least one test image, j, by filtering said at least one test image, j, with each of said at least two adaptive high-pass filters, HF, wherein said extracting is performed by lowering the threshold of strength, t, wherein t=t−

c, wherein c is a user-selected constant based on application-specific conditions;

selecting, in each pass through each of said at least two adaptive high-pass filters, HF, at least one keypoint, KP, in said at least one test image, j;

determining, in each pass through each of said at least two adaptive high-pass filters, HF, whether the number of keypoints, KP, is greater than or equal to the number of said descriptors, N, wherein said determination is expressed as KP≧

N;

when, in each pass through each of said at least two adaptive high-pass filters, said number of keypoints, KP, is not greater than or equal to the number of said at least one SIFT descriptor, N, iterating through said extracting a plurality of keypoints, KP, from said at least one test image, j, wherein said extracting is performed by lowering the threshold of strength, t, wherein t=t−

c, wherein c is a user-selected constant based on application-specific conditions, until said number of keypoints KP is greater than or equal to the number of said at least one SIFT descriptor, N;

extracting, in each pass through each of said at least two adaptive high-pass filters, HF, said at least one SIFT descriptor, N, from said at least one test image, j; and

summing said at least one SIFT descriptor, N, from each pass through each of said at least two adaptive high-pass filters, HF, to yield a set of N-SIFT descriptors.

10. The method according to claim 9, wherein said recognition task, comprises:

solving a matrix, X, by using a sparse solution, wherein said matrix X is said sparse solution for each of said at least one SIFT descriptor N in columns of matrix Y, wherein Y=AX; and

summing number of matches per class, wherein said number of matches per class is vector q, wherein said vector q is the vector solution extracted from said matrix X for each class, c.

11. The method according to claim 10, wherein said determination whether to learn task, comprises:

initializing model parameters for said summed number of matches per class, q;

initializing the number of clusters to a user defined number, K, wherein each of said cluster is a Gaussian distribution selected from the group of matches consisting of said summed number of matches per class vector, q;

initializing the number of iterations, M, to a user selected number to update said model parameters and each of said Gaussian distribution, and initializing the actual number of iterations performed, N_iter, to 1 on the first iteration of said number of iterations, M, wherein N_iter increases by 1 each time through said iteration;

updating model parameters for each potential k cluster;

determining whether the actual number of iterations performed, N_iter, is greater than M;

when the number of actual iterations performed, N_iter, is not greater than M, iterating through said updating model parameters for each potential k cluster task, and said determining whether the actual number of iterations performed, N_iter, is greater than M, until the actual number of iterations performed, N_iter, is greater than M;

finding a cluster with the highest mean value, μ

_c;

when said cluster with the highest mean value, μ

_c, does not belong in the same class as q_max, learning does not occur and said dictionary A_c(j-1) is not updated;

when said cluster with the highest mean value, μ

_c, belongs in the same class as q_max, determine whether said cluster with the highest mean value, μ

_c, has a single member;

wherein when said cluster with the highest mean value, μ

_c, does not have a single member, learning does not occur and said dictionary A is not updated; and

wherein when said cluster with the highest mean value, μ

_c, does have a single member, learning does occur and said dictionary A_c(j-1), is updated to A_cj, with the columns of matrix Y.

12. The method according to claim 11, wherein said adapting said dictionary, A, using reversed-orthogonal matching pursuit (OMP), comprises:

inputting said columns of updated matrix, Y_j, for said dictionary, A_c(j-1), and q_max into a matrix X_j by using a sparse solution, wherein said matrix X_j is said sparse solution, wherein A_c(j-1)=Y_jX_j, wherein Y_j is the set of N-SIFT descriptors extracted from said at least one test image, j;

finding a complement of X_j, wherein Y_j=Y_j(X_j)∪

Y_j(X_j^c), wherein X_j^c is the set of N-SIFT descriptors not used in said sparse solution; and

updating said dictionary A_c(j-1) to A_c(j), wherein A_cj=[A_c(j-1) Y(X_j^c)].

13. The method according to claim 12, wherein said outputting a class having the maximum output, q_max, comprises:

normalizing said number of matches per class, vector q, to find said class with the maximum energy, q_max; and

outputting said class having maximum energy, q_max.

14. The method according to claim 8, wherein said tangible medium is a visual display screen.