|
|
Facial Age Classification for Unconstrained Face Images |
|
|
|
1. Introduction
l Recently, automatic age
classification using face information has gained an increasing interest due to
its emerging applications such as
ž Forensic art
ž Electronic customer
relationship management (ECRM)
ž Security control and
surveillance monitoring
ž Biometrics
l Age classification in the real
world is a very challenging task due to the large variation of face appearance.
ž A variety of human races,
genders, facial expression, pose etc.
l Most of existing age classification
methods globally modeled faces of age groups (i.e., representing each age group
as a single model).
ž Globally modeling age groups
was problematic in an uncontrolled condition
ž This is because faces of same
age group could be significantly different or faces in different age groups
could be similar due to the large face appearance variations
l In this research, we propose a
new age classification method suitable for dealing with large face appearance
variations
ž We decompose the whole
training faces of an age group into a set of face clusters.
ž Using a few nearest face
clusters of each age group, we can avoid the classification degradation due to
some undesirable faces
ž To measure the closeness
between a test face and a face cluster, we compute the distance between the
test face and the face cluster centroid that represents the representative face
appearance.
ž To further reflect the
variation of face appearance, we compute the distance between a test face and
the face sample distribution
2. Age Classification Using Local Age Group Modeling
Fig. 1. Overview of the proposed age
classification based on local age group modeling.
l Training stage
ž Feature extraction: captures
rich age-related information from face image
ž Face clustering: groups faces
with similar appearance
ž Construction of local age
group model: Each face cluster is converted to a local age group model that
enables matching with test face image in the test stage
l Test stage
ž Selecting nearest local age
group models: selects the few local age group models nearest to the test face,
in each age group
ž Computation of similarity
scores: obtains overall similarity scores for each age group, by using the few
local age group models
ž Classification: determines the
age group label by finding the highest similarity score among all age groups
2.1. Construction of Local Age Group Models from Training Faces
l Step 1: We construct training set 
of the i-th age
group, which consists of 
feature vectors denoted by 
.
ž For feature extraction, we use
Local Binary Pattern (LBP) histogram followed by Fisher¡¯s linear discriminant
analysis (FLDA).
l Step 2: Using face clustering, we divide 
into a set of face clusters where each
has similar characteristics (in terms of gender, races, facial pose etc).
ž Because the number of face
clusters may vary with different datasets, hierarchical agglomerative
clustering (HAC) is used instead of kmeans clustering.
ž The set of face clusters for
the i-th age group is defined as 
where 
is the k-th
face cluster in 
.
l Step 3: We construct the local models of .
ž Compute cluster centroid: to encode
representative face appearance of cluster
ž Compute linear subspace: to
encode sample variation information of cluster
2.2. Classification of Test Face by Using Local Age Group Models
l Step 1: Compute the closeness between the test sample 
and local age group model
.
ž Distance between and cluster centroid 
of 
:
|
|
(1) |
ž Distance between and linear subspace 
of 
:
(This is equivalent to the PCA
(Principal Component Analysis) residual between and the reconstruction of projected onto the linear subspace 
)
|
|
(2) |
ž In (2), 
corresponds to the reconstruction of (subtracted by the cluster centroid)
l Step 2: Convert the distances 
and 
into the similarity scores 
and 
.
l Step 3: Compute the combined similarity score 
by weighted sum of and .
l Step 4: Compute the overall similarity score 
for each age group, as the average of the
R largest similarity scores among all 
where 
.
ž R is too small: the
classification may overfit the training data
ž R is too large: degrade the
classification performance by incorporating the training face images that are
very different from the test face images
l Step 5: The age group label () of the feature vector is determined by finding the maximum
similarity score among G age groups
|
|
(3) |
|
|
|
|
(a) |
(b) |
Fig. 2. Conceptual
illustration to show the usefulness of computing closeness between test face
and local age group model by considering variation of face appearance.
(a) the distances between the query face and the centroids of
two clusters of different age groups. The distances are similar in this
example.
(b) the distances between the query face and the linear
subspaces of two clusters. The distances are more different, which is useful
for classification.
3. Experiment
<Experimental setup>
l
Used database: Images of Groups database (DB)
ž
This DB contains 28,231 faces from 5,080 Flickr images
ž
Faces often had low-resolution, occlusion, dark
glasses, unusual facial expressions etc. (see Fig. 3)
ž
We cropped the face regions and normalized them to 60
(h) x 48 (w) pixel images using eye center positions.
l
Construction of dataset 1
ž
Training set: Randomly selected 3,500 (500 per age
group) face images
ž
Test set: The remaining 1,050 (150 per age group) face
images
ž
This dataset 1 was termed as ¡®Full set¡¯
l
Construction of dataset 2
ž
To study age classification on faces with reasonable
resolution, we collected 12,080 face images with the eye distance more than 24
pixels
ž
Training set: Randomly selected 2,080 (300 per age
group except for the age group ¡®8-12¡¯ where only 280 face images were available)
face images
ž
Test set: The remaining 1,050 (100 per age group
except for the age group ¡®8-12¡¯ where only 64 face images were available) face
images
ž
This dataset 2 was termed as ¡®More than 24 pixels¡¯
Fig. 3. Example face images from the
Images of Groups database.
<Experimental results>
l
Comparison with state-of-the-arts
ž
A significant improvement over the baseline can be
made about 8%.
ž
This result reveals that the local age modeling could
be advantageous over the global age modeling in the presence of huge variation
in face appearance.
ž
Our proposed method outperforms the other
state-of-the-arts in Table 1.
l
Analysis for confusion matrix
ž
Table 2 shows the confusion matrix for the proposed
method.
ž
For the age group ¡®0-2¡¯ and ¡®66+¡¯, the proposed method
achieves acceptable classification accuracies (around 75%).
ž
Recognizing the other age groups are relatively
difficult as these age groups could be confused with both the younger and elder
age groups.
Table 2. Comparisons with the baseline and the state-of-the-arts.
<References for comparison methods>
* Baseline: The proposed method in which face clustering was not applied
while the other conditions such as feature extraction and classification
remained the same as the proposed method.
[1] A. Gallagher and T. Chen, ¡°Understanding Images of
Groups of People,¡± IEEE International Conference on Computer Vision and Pattern
Recognition (CVPR), 2009.
[2] C. Shan, ¡°Learning Local Features for Age
Estimation on Real-life Faces,¡± ACM International Workshop on Multimodal Pervasize Video Analysis (MPVA), 2010.
[3] J. Ylioinas, A. Hadid, and M. Pietikainen, ¡°Age
Classification in Unconstrained Conditions Using LBP Variants,¡± IEEE
International Conference on Pattern Recognition, 2012.
[4] J.-G. Wang, W.-Y. Yau, and H. L. Wang, ¡°Age
Categorization via ECOC with Fused Gabor and LBP Features,¡± IEEE Workshop Applic. Comput. Vis. 2012.
Table 2. Confusion matrix using the proposed method (For full set).
Row and column mean actual and predicted age groups.
4. Conclusions
l
To reduce the effect of the factors unrelated to age,
only a few important local models of each age group affected the age
classification.
l
Experimental results using a real-world aging database
containing images from Flickr verified the effectiveness of the proposed method
using local age group modeling.
* Contact Person: Prof. Yong Man Ro (ymro@kaist.ac.kr)
1. S. H. Lee and Y. M. Ro, ¡°Local Age Group Modeling in Unconstrained Face Images for Facial Age Classification,¡± IEEE Conference on Image Processing, 2014.