Automatic face recognition
Automatic face recognition. In a first example embodiment, a method for automatic face recognition includes several acts. First, a face pattern and two eye patterns are detected. Then, the face pattern is normalized. Next, the normalized face pattern is transformed into a normalized face feature vector of Gabor feature representations. Then, a difference image vector is calculated. Next, the difference image vector is projected to a lower-dimensional intra-subject subspace extracted from a pre-collected training face database. Then, a square function is applied to each component of the projection. Next, a weighted summation of the squared projection is calculated. Then, the previous four acts are repeated for each normalized gallery image feature vector. Finally, the face pattern in the probe digital image is classified as belonging to the gallery image with the highest calculated weighted summation where the highest calculated weighted summation is above a predefined threshold.
1. A method for automatic face recognition, the method comprising the following acts:
i) detecting a face pattern in a probe digital image;
ii) detecting two eye patterns in the detected face pattern;
iii) normalizing the face pattern by rotating and resizing the face pattern to place centers of the two eye patterns on predetermined pixels;
iv) transforming the normalized face pattern into a normalized face feature vector of Gabor feature representations;
v) calculating a difference image vector using the normalized face feature vector and a normalized gallery image feature vector of a gallery image in a gallery of images;
vi) projecting the difference image vector to a lower-dimensional intra-subject subspace extracted from a pre-collected training face database;
vii) applying a square function to each component of the projection;
viii) calculating a weighted summation of the squared projection;
ix) repeating acts v)-viii) for each normalized gallery image feature vector in the gallery of images; and
x) classifying the face pattern in the probe digital image as belonging to the gallery image with the highest calculated weighted summation where the highest calculated weighted summation is above a predefined threshold.
2. The method as recited in claim 1 , wherein the act iii) further comprises applying a mask to the face pattern to remove non-face portions.
3. The method as recited in claim 2 , wherein the act iii) further comprises performing histogram equalization on the masked face pattern.
4. The method as recited in claim 1 , wherein the act iv) comprises transforming the normalized face pattern I(x) into a normalized face feature vector of Gabor feature representations G uv (x), where each Gabor feature representation G uv (x) is obtained according to the following equation:
G uv ( x )=( I*φ uv )( x )
where:
x=[x, y] denotes the x- and y-coordinates of pixels of the normalized face pattern I(x); and
φ u,v is a Gabor wavelet filter defined as follows:
φ
u
,
v
(
x
)
=
k
uv
2
ω
2
exp
(
-
k
uv
2
x
2
2
ω
2
)
[
exp
(
ⅈ
k
uv
x
)
-
exp
(
-
ω
2
2
)
]
where:
u, v define the orientations and scales of a Gabor wavelet kernel k;
ω is a sigma of a Gaussian envelope;
∥.∥ denotes a norm operator; and
k uv is defined as
k
uv
=
k
max
f
exp
(
ⅈ
π
u
8
)
where k max is a maximum frequency and f is a spacing factor between kernels in a frequency domain.
5. The method as recited in claim 4 , wherein:
k
max
=
π
2
;
f
=
2
;
ω
=
2
π
;
u
=
{
0
,
…
,
7
}
;
and
v
=
{
0
,
…
,
4
}
.
6. The method as recited in claim 1 , wherein the act vi) comprises projecting the difference image vector δ i to a lower-dimensional intra-subject subspace extracted from a pre-collected training face database resulting in a projection {circumflex over (δ)} that is calculated according to the following formula:
{circumflex over (δ)} i =A T δ i
where A T is a transformation matrix.
7. The method as recited in claim 6 , wherein the act vii) comprises applying a square function to each component of the projection resulting in the squared projection {tilde over (δ)} i that is calculated according to the following formula:
{tilde over (δ)} i =[{tilde over (δ)} 1 i , . . . ,{tilde over (δ)} M i ] T =[({circumflex over (δ)} 1 i ) 2 , . . . ,({circumflex over (δ)} M i ) 2 ] T
where M components are firstly partitioned into T groups sequentially with each group containing L components.
8. The method as recited in claim 7 , wherein the act viii) comprises calculating the weighted summation S of the squared projection {tilde over (δ)} i according to the following formulas:
S
=
∑
j
=
1
M
w
j
δ
~
j
=
∑
m
=
1
T
∑
n
=
1
L
w
^
m
δ
~
(
m
-
1
)
L
+
n
=
∑
m
=
1
T
w
^
m
(
∑
n
=
1
L
δ
~
(
m
-
1
)
L
+
n
)
=
∑
m
=
1
T
w
^
m
s
m
=
w
^
T
s
where
:
s
m
=
∑
n
=
1
L
δ
~
(
m
-
1
)
L
+
n
;
w
^
=
[
w
^
1
,
…
,
w
^
T
]
T
;
s
=
[
s
1
,
…
,
s
T
]
T
;
and
w
=
[
[
w
^
1
,
…
,
w
^
1
]
,
[
w
^
2
,
…
,
w
^
2
]
,
…
,
[
w
^
T
,
…
,
w
^
T
]
]
T
denotes a regularized weight vector.
9. A method for automatic face recognition, the method comprising the following acts:
i) obtaining a normal to a separation hyperplane ŵ=[ŵ 1 , . . . ,ŵ T ] through SVM training;
ii) normalizing G gallery images and transforming the normalized G gallery images into corresponding Gabor feature vectors, denoted as g i , i=1, . . . , G;
iii) normalizing a probe image and transforming the normalized probe image into a corresponding Gabor feature vector, denoted as p;
iv) calculating an image difference vector δ, denoted as δ=p−g i ;
v) projecting the image difference vector δ to intra-subject space A T , thereby obtaining a projection {circumflex over (δ)}=A T δ;
vi) applying a square function to the projection {circumflex over (δ)} to obtain {tilde over (δ)};
vii) M components of {tilde over (δ)} are firstly partitioned into T groups sequentially with each group containing L components with {tilde over (δ)}=[[{tilde over (δ)} 1 , . . . ,{tilde over (δ)} L ],[{tilde over (δ)} L+1 , . . . ,{tilde over (δ)} 2L ], . . . ,[{tilde over (δ)} (T−1)L+1 , . . . ,{tilde over (δ)} TL ]] T ;
viii) calculating a weighted summation S i according to the following equation:
S
=
∑
j
=
1
M
w
j
δ
~
j
=
∑
m
=
1
T
∑
n
=
1
L
w
^
m
δ
~
(
m
-
1
)
L
+
n
=
∑
m
=
1
T
w
^
m
(
∑
n
=
1
L
δ
~
(
m
-
1
)
L
+
n
)
=
∑
m
=
1
T
w
^
m
s
m
=
w
^
T
s
;
ix) for each g i , repeating acts iv)-viii); and
x) determining that the probe image belongs to the gallery subject associated with the highest S i value where the highest S i value is above a predefined threshold,
where:
s
m
=
∑
n
=
1
L
δ
~
(
m
-
1
)
L
+
n
;
w
^
=
[
w
^
1
,
…
,
w
^
T
]
T
;
s
=
[
s
1
,
…
,
s
T
]
T
;
w
=
[
[
w
^
1
,
…
,
w
^
1
]
,
[
w
^
2
,
…
,
w
^
2
]
,
…
,
[
w
^
T
,
…
,
w
^
T
]
]
T
denotes a regularized weight vector.
10. The method as recited in claim 9 , wherein the act vi) comprises applying a square function to obtain {tilde over (δ)} according to the following equation:
{tilde over (δ)} i =[{tilde over (δ)} 1 i , . . . ,{tilde over (δ)} M i ] T =[{circumflex over (δ)} 1 i ) 2 , . . . ,({circumflex over (δ)} M i ) 2 ] T .
11. The method as recited in claim 9 , wherein the normalizing acts ii) and iii) each comprise:
detecting a face pattern in the respective image;
detecting two eye patterns in the detected face pattern;
rotating and resizing the face pattern to place centers of the two eye patterns on predetermined pixels;
applying a mask to the face pattern to remove non-face portions; and
performing histogram equalization on the masked face pattern.
12. One or more non-transitory computer-readable media having computer-readable instructions thereon which, when executed by a programmable processor, implement a method for automatic face recognition, the method comprising the acts of:
i) detecting a face pattern in a probe digital image;
ii) detecting two eye patterns in the detected face pattern;
iii) normalizing the face pattern by rotating and resizing the face pattern to place centers of the two eye patterns on predetermined pixels;
iv) transforming the normalized face pattern into a normalized face feature vector of Gabor feature representations;
v) calculating a difference image vector using the normalized face feature vector and a normalized gallery image feature vector of a gallery image in a gallery of images;
vi) projecting the difference image vector to a lower-dimensional intra-subject subspace extracted from a pre-collected training face database;
vii) applying a square function to each component of the projection;
viii) calculating a weighted summation of the squared projection;
ix) repeating acts v)-viii) for each normalized gallery image feature vector in the gallery of images; and
x) classifying the face pattern in the probe digital image as belonging to the gallery image with the highest calculated weighted summation where the highest calculated weighted summation is above a predefined threshold.
13. The one or more non-transitory computer-readable media as recited in claim 12 , wherein the act iii) further comprises applying a mask to the face pattern to remove non-face portions and performing histogram equalization on the masked face pattern.
14. The one or more non-transitory computer-readable media as recited in claim 12 , wherein the act iv) comprises transforming the normalized face pattern I(x) into a normalized face feature vector of Gabor feature representations G uv (x), where each Gabor feature representation G uv (x) is obtained according to the following equation:
G uv ( x )=( I*φ uv )( x )
where:
x=[x,y] denotes the x- and y-coordinates of pixels of the normalized face pattern I(x);
φ u,v is a Gabor wavelet filter defined as follows:
φ
u
,
v
(
x
)
=
k
uv
2
ω
2
exp
(
-
k
uv
2
x
2
2
ω
2
)
[
exp
(
ⅈ
k
uv
x
)
-
exp
(
-
ω
2
2
)
]
where:
u, v define the orientations and scales of a Gabor wavelet kernel k;
ω is a sigma of a Gaussian envelope;
∥.∥ denotes the norm operator; and
k uv is defined as
k
uv
=
k
max
f
exp
(
ⅈ
π
u
8
)
where k max is a maximum frequency and f is a spacing factor between kernels in a frequency domain.
15. The one or more non-transitory computer-readable media as recited in claim 14 , wherein:
k
max
=
π
2
;
f
=
2
;
ω
=
2
π
;
u
=
{
0
,
…
,
7
}
;
and
v
=
{
0
,
…
,
4
}
.
16. The one or more non-transitory computer-readable media as recited in claim 12 , wherein the act vi) comprises projecting the difference image vector δ i to a lower-dimensional intra-subject subspace extracted from a pre-collected training face database resulting in a projection {circumflex over (δ)} that is calculated according to the following formula:
{circumflex over (δ)} i =A T {circumflex over (δ)} i
where A T is defined as a transformation matrix.
17. The one or more non-transitory computer-readable media as recited in claim 16 , wherein the act vii) comprises applying a square function to each component of the projection resulting in the squared projection {tilde over (δ)} i that is calculated according to the following formula:
{tilde over (δ)} i =[{tilde over (δ)} 1 i , . . . , {tilde over (δ)} M i ] T =[({circumflex over (δ)} 1 i ) 2 , . . . ,({circumflex over (δ)} M i ) 2 ] T
where M components are firstly partitioned into T groups sequentially with each group containing L components.
18. The one or more non-transitory computer-readable media as recited in claim 17 , wherein the act viii) comprises calculating the weighted summation S of the squared projection {tilde over (δ)} i according to the following formulas:
S
=
∑
j
=
1
M
w
j
δ
~
j
=
∑
m
=
1
T
∑
n
=
1
L
w
^
m
δ
~
(
m
-
1
)
L
+
n
=
∑
m
=
1
T
w
^
m
(
∑
n
=
1
L
δ
~
(
m
-
1
)
L
+
n
)
=
∑
m
=
1
T
w
^
m
s
m
=
w
^
T
s
where
:
s
m
=
∑
n
=
1
L
δ
~
(
m
-
1
)
L
+
n
;
w
^
=
[
w
^
1
,
…
,
w
^
T
]
T
;
s
=
[
s
1
,
…
,
s
T
]
T
;
w
=
[
[
w
^
1
,
…
,
w
^
1
]
,
[
w
^
2
,
…
,
w
^
2
]
,
…
,
[
w
^
T
,
…
,
w
^
T
]
]
T
denotes a regularized weight vector.
19. An image processing apparatus comprising:
an electronic display; and
a processor in electronic communication with the electronic display; and
one or more computer-readable media in electronic communication with the processor, the one or more computer-readable media having computer-readable instructions thereon which, when executed by the processor, cause the processor to:
i) detect a face pattern in a probe digital image;
ii) detect two eye patterns in the detected face pattern;
iii) normalize the face pattern by rotating and resizing the face pattern to place centers of the two eye patterns on predetermined pixels;
iv) transform the normalized face pattern into a normalized face feature vector of Gabor feature representations;
v) calculate a difference image vector using the normalized face feature vector and a normalized gallery image feature vector of a gallery image in a gallery of images;
vi) project the difference image vector to a lower-dimensional intra-subject subspace extracted from a pre-collected training face database;
vii) apply a square function to each component of the projection;
viii) calculate a weighted summation of the squared projection;
ix) repeat acts v)-viii) for each normalized gallery image feature vector in the gallery of images;
x) classify the face pattern in the probe digital image as belonging to the gallery image with the highest calculated weighted summation where the highest calculated weighted summation is above a predefined threshold; and
xi) send data corresponding to the gallery image to which the probe digital image belongs to the electronic display for display thereon.
20. The image processing apparatus as recited in claim 19 , wherein:
the image processing apparatus comprises a photo viewer;
the one or more computer-readable media comprises one or more of a RAM, a ROM, and a flash EEPROM; and
the electronic visual display comprises a liquid crystal display.