# US 2010/0166266 A1 · Method for Identifying Faces in Images with Improved Accuracy Using Compressed Feature Vectors

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2010/0166266 A1
- Document ID: US20100166266A1
- Status: Published application
- Kind: A1
- Application number: 12/345,919
- Filed: 2008-12-30
- Published: 2010-07-01
- Earliest claimed: 2008-12-30
- Assignees: Mitsubishi Electric Research Laboratories, Inc.
- Inventors: Michael Jeffrey Jones; Jay Thornton
- Web page: https://patentgrep.com/patent/US20100166266A1

## CPC

- G06V 40/172: {Classification, e.g. identification}
- G06F 18/2321: using statistics or function optimisation, e.g. modelling of probability density functions
- G06F 18/2413: based on distances to training or reference patterns
- G06V 10/446: {using Haar-like filters, e.g. using integral image techniques}
- G06V 10/763: {Non-hierarchical techniques, e.g. based on statistics of modelling distributions}
- G06V 10/764: using classification, e.g. of video objects
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

A method identifies an unknown face in an input image using reference images of known faces. A Haar-like feature vector is extracted from each image. The vectors are compressed. An L1 norm is determined between the compressed feature vector of the input image and each compressed feature vector from the set of reference images to determine a most similar reference image. The identity of the face associated with the most similar reference image is assigned as the identity of the unknown face in the input image.

## Claims

### Claim 1

A method for comparing an input image of an unknown face with a reference image of a known face, comprising the steps of:
extracting a feature vector from each image;
compressing the feature vectors into compressed feature vectors; and
comparing the compressed feature vectors to determine a similarity score for the two faces.

### Claim 2 (depends on 1)

The method of claim 1, in which the features are rectangular Haar-like features defined as a weighted sum or difference of sums of pixel values within two or more rectangular patches of pixels in the images, and wherein the rectangular patches are contiguous at any position and scale within the images.

### Claim 3 (depends on 2)

The method of claim 2, wherein the Haar-like features are determined from integral images.

### Claim 4 (depends on 1)

The method of claim 1, wherein the feature vector is extracted and compressed from each of a set of reference images of a set of corresponding known faces, and the compressed feature vector of the input image is compared with each feature vector of the set of reference images to determine a similarity score for the unknown face with each known face.

### Claim 5 (depends on 1)

The method of claim 1, in which n is a number of features in each feature vector, ν=[f1, f2, . . . , fn] is the feature vector extracted from each image, fi is an ith feature of the feature vector 84 x where 1≦i≦n, 1=[μ1, μ2, . . . , μn] is a vector of means of the features, σ=[σ1, σ2, . . . , σn] is a vector of standard deviations of the features, s is a scalar constant controlling a width of a truncation region about μ, and the compressing further comprises:
determining a value Fi for each extracted feature value fi according to
T
i
=
2
b
-
1
×
(
1
+
(
f
i
-
μ
i
s


σ
i
)
)
,

C
i
=
{
2
b
-
1
for


T
i
>
2
b
-
1
T
i
for


0
≤
T
i
≤
2
b
-
1
,
0
for


T
i
<
0


F
i
=
truncate

(
C
i
)
.
where b is a number of bits used to store the value Fi;

### Claim 6 (depends on 5)

The method of claim 5, wherein the mean and standard deviation is determined from a set of training images, wherein a total number of training images is r, and n is a total number of features in each feature vector, and wherein a feature vector extracted from an xth training image is νx=[f1, f2, . . . , fn] where 1≦x≦r, and an ith feature value of νx is fi where 1≦i≦n, and a vector of means μ=[μ1, μ2, . . . , μn] is determined according to
μ
i
=
∑
x
=
1
r

(
f
i
)
x
r
,
1
≤
i
≤
n
,
and a vector of standard deviations σ=[σ1, σ2, . . . , σn] is determined according to
σ
i
=
∑
x
=
1
r

(
(
f
i
)
x
-
μ
i
)
2
r
,
1
≤
i
≤
n
.

### Claim 7 (depends on 1)

The method of claim 1, in which F=[F1, F2, . . . , Fn] is the compressed feature vectors of the input image, G=[G1, G2, . . . , Gn] is the compressed feature vectors of the reference image, M is a maximum value of a L1 norm of the compressed feature vectors F and G, and the similarity score is determined according to
Similarity


(
F
,
G
)
=
M
-
∑
i
=
1
n


F
i
-
G
i

.

### Claim 8 (depends on 1)

The method of claim 1, further comprising:
normalizing the feature vectors to produce normalized feature vectors; and
quantizing the normalized feature vectors to produce quantized feature vectors, and wherein the comparing uses the quantized feature vectors.

### Claim 9 (depends on 4)

The method of claim 4, wherein each reference image has an associated identity, and the associated identity of the reference image with a largest similarity scored when compared to the input face as long as the largest similarity score is above a predetermined threshold.

### Claim 10 (depends on 5)

The method of claim 5, wherein b is eight.

### Claim 11 (depends on 5)

The method of claim 5, wherein b is one.

### Claim 12 (depends on 5)

The method of claim 5, wherein the standard deviation is in a form of sσ, wherein s is a real number.