L1-optimized AAM alignment
An Active Appearance Model, AAM, uses an L 1 minimization-based approach to aligning an input test image. In each iterative application of its statistical model fitting function, a shape parameter coefficient p and an appearance parameter coefficient λ within the statistical model fitting function are updated by L 1 minimization. The AAM further includes a canonical classifier to determine if an aligned image is a true example of the class of object being sought before the AAM is permitted to output its aligned image.
1. An Active Appearance Model machine, comprising:
a learn module providing a model image of a class of object, said model image being created by combining feature information from an image library of true image samples of said class of object, said learn module further providing a statistical model fitting function defining shape and appearance features of said class of object;
an input for receiving an input image; and
an align module optimizing said statistical model fitting function to determine a best fit of said model image and said input image through iterative applications of said statistical model fitting function to produce an aligned image;
wherein in each of said iterative applications, a shape parameter coefficient p and an appearance parameter coefficient λ within said statistical model fitting function are updated by L 1 minimization, and said L 1 minimization being defined as:
min
Δ
p
,
Δ
λ
A
λ
+
[
SD
A
]
C
(
λ
)
[
Δ
p
Δ
λ
]
Δ
p
-
I
(
p
)
|
1
wherein A=appearance base (i.e. current appearance of the model image within a current iteration), SD=a steepest descent matrix for shape coefficients, C(λ)=coefficient matrix dependent on current appearance parameters, Δp=update in the shape projection coefficients, Δλ=update in the appearance projection coefficients, and I(p)=appearance extracted from the input image based on current shape parameters within a current iteration.
2. The Active Appearance Model machine of claim 1 , wherein the appearance basis A is comprised of eigenvectors, and A is also computed by L 1 -decomposition.
3. The Active Appearance Model machine of claim 1 , wherein said class of object is a human face class.
4. The Active Appearance Model machine of claim 1 further including:
canonical class classifier to determine if said aligned image is a true representation of said class of object; and
an output for outputting said fitted image only if said canonical class classifier determines that said fitted image is a true representation of said class of object.
5. The Active Appearance Model machine of claim 4 , wherein said canonical classifier has access to a library of fitted images (i.e. a library of previously aligned faces), and is trained using said library of fitted images, said library of fitted images being comprised of a positive training set of true fitted images and a negative training set of untrue fitted images.
6. The Active Appearance Model machine of claim 5 , wherein said positive training set is comprised of manually verified true images output from said align module as a result of submitting said image library of true image samples of said class of object to said align module.
7. The Active Appearance Module machine of claim 6 , wherein said negative training set is comprised of object images from said positive training set to which perturbations have been introduced, wherein said perturbations do not affect the size or perimeter outline of said object images.
8. The Active Appearance Model machine of claim 5 , wherein the fitted images within said positive training set and negative training set have characteristic features of said class of object manually identified and demarcated.
9. The Active appearance Model machine of claim 8 , wherein the size and perimeter outline of the fitted images within said positive training set and negative training set is the same as the size and perimeter outline of said model image.
10. The Active appearance Model machine of claim 5 wherein the untrue fitted images are constructed from the true fitted images by introducing perturbations related to eigenvalues within said statistical model.
11. The Active appearance Model machine of claim 5 wherein the untrue fitted images are constructed from the true fitted images by applying pre-defined translation, scale, and rotation offsets to the true fitted images, said scale offsets not affecting the size or perimeter outline of the true fitted images.
12. The Active Appearance Model machine of claim 5 , wherein the number of untrue fitted images in said negative training set is greater than then number of true fitted images in said positive training set.
13. The Active Appearance Model machine of claim 12 , wherein there is a 10:1 ratio in the number of images in said negative training set to the number of images in said positive training set.
14. An Active Appearance Model machine, comprising:
a learn module providing a model image of a class of object, said model image being created by combining feature information from an image library of true image samples of said class of object, said learn module further providing a statistical model fitting function defining shape and appearance features of said class of object;
an input for receiving an input image;
an align module optimizing said statistical model fitting function to determine a best fit of said model image and said input image through iterative applications of said statistical model fitting function to produce an aligned image;
canonical class classifier to determine if said aligned image is a true representation of said class of object; and
an output for outputting said aligned image only if said canonical class classifier determines that said aligned image is a true representation of said class of object.
15. The Active Appearance Model machine of claim 14 , wherein said canonical classifier has access to a library of fitted images (i.e. a library of previously aligned faces), and is trained using said library of fitted images, said library of fitted images comprised of a positive training set of true fitted images and a negative training set of untrue fitted images.
16. The Active Appearance Model machine of claim 15 , wherein said positive training set is comprised of manually verified true images output from said align module as a result of submitting said image library of true image samples of said class of object to said align module.
17. The Active Appearance Module machine of claim 16 , wherein said negative training set is comprised of object images from said positive training set to which perturbations have been introduced, wherein said perturbations do not affect the size or perimeter outline of said object images.
18. The Active Appearance Model machine of claim 15 , wherein the fitted images within said positive training set and negative training set have characteristic features of said class of object manually identified and demarcated.
19. The Active Appearance Model machine of claim 18 , wherein the size and perimeter outline of the fitted images within said positive training set and negative training set are the same as the size and perimeter outline of said model image.
20. The Active Appearance Model machine of claim 15 wherein the untrue fitted images are constructed from the true fitted images by introducing perturbations related to eigenvalues within said statistical model.
21. The Active Appearance Model machine of claim 15 wherein the untrue fitted images are constructed from the true fitted images by applying pre-defined translation, scale, and rotation offsets to the true fitted images, said scale offsets not affecting the size or perimeter outline of the true fitted images.
22. The Active Appearance Model machine of claim 15 , wherein there is a 10:1 ratio in the number of images in said negative training set to the number of images in said positive training set.
23. The Active Appearance Model machine of claim 14 , wherein:
said align module applies L 1 minimization to compute parameter updates for said statistical model fitting function; and
in each of said iterative applications, a shape parameter coefficient p and an appearance parameter coefficient λ are updated, and said L 1 minimization is defined as:
min
Δ
p
,
Δ
λ
A
λ
+
[
SD
A
]
C
(
λ
)
[
Δ
p
Δ
λ
]
Δ
p
-
I
(
p
)
|
1
wherein A=appearance base (i.e. current appearance of the model image within a current iteration); SD=a steepest descent matrix for shape coefficients; C(λ)=coefficient matrix dependent on current appearance parameters; Δp=update in the shape projection coefficients; Δλ=update in the appearance projection coefficients; and I(p)=appearance extracted from the input image based on current shape parameters.
24. The Active Appearance Model machine of claim 23 , wherein the appearance basis A is computed by L 1 -decomposition.
25. The Active Appearance Model machine of claim 14 , wherein said class of object is a human face.