Академический Документы
Профессиональный Документы
Культура Документы
Introduction: Pose-invariant gender classication is one of the most difcult and challenging tasks in computer vision. Generally, gender classication encompasses two main steps: feature extraction and
classication. The method used in feature extraction should give adequate features, otherwise it will be misclassied. Hence, there are
many methods to extract features that robustly recognise gender under
the restriction of a one two-dimensional (2D) single training sample
for each two-class. The common gender classication approaches are
mostly categorised into two separate types: geometric-based and featurebased methods. Geometric features contain information about the
location and shape of facial features. Geometric features are more sensitive to noise and tracking errors than appearance features.
The feature-based approaches [13] utilise image lters; either on the
whole faces to construct global features, or several particular face
regions to construct local features, to extract the feature changes in the
face image. Popular approaches are Gabor wavelets [4], the local binary
patterns (LBP) operator [2], the interlaced derivative patterns (IDP) [3],
the local derivative pattern [5] and so on. Recently, Shan proposed
gender recognition on real-world faces [3]. In this method, LBP was
employed to describe faces, and Adaboost was used to select the discriminative LBP features. Finally, gender classication results were acquired
by applying the support vector machine (SVM) with the boosted LBP features. Moreover, Shobeirinejad and Gao [3] presented the IDP method as
a novel feature extraction, which is a derivative-based technique to extract
discriminative facial features for gender classication. This method operates on a neighbourhood around a pixel and concatenates the extracted
regional feature distributions as a feature vector.
In this Letter, a new approach is proposed for handling face poses in
gender classication based on 3D reconstruction and synthesising of
the face. Accordingly, a 3D model was initially reconstructed from 2D
frontal face images with facial expression. To reconstruct a 3D model
from each human frontal face with facial expression, a facial expression
generic elastic model (FE-GEM) was used. Then, each 3D reconstructed
face in the gallery was synthesised to all possible views and a feature
library matrix (FLM) was generated based on the yaw angles of face
poses for each two-class (male and female). On the other hand, automatic
head pose estimation by the eye-based normalisation approach was used
to extract the yaw angles of face poses. Therefore, for each two-class, an
array of FLM was selected based on the yaw angles of face poses which
were estimated from automatic head pose estimation. Finally, gender
classication was performed by the SVM [6] between the FLMs selected
arrays and the target images features.
Constructing a 3D database: To reconstruct the 3D face model from a
single image, the FE-GEM method [7] is used. Moeini et al. [7] proposed a method for 3D face modelling that is robust to facial expression.
In this method, three FE-GEMs were employed in the GEM framework
to generate three models for the input image. Then, these three generated
models were combined based on the similarity distance around the lips
and mouth of input images with the corresponding distance in facial
expression generic models. Finally, the desired model was generated
for input images from a mixture of the combined models. To construct
the 3D database for each gender, one 2D frontal image was registered
from each person with an arbitrary facial expression. Therefore, 2D
yaw angle (Y )
90
+90
+90
yaw angle (Y )
PCA
low-dimensional
feature
train images
offline
3D face
database
3D face
reconstruction
yaw
pose matrix P i,Y
A i,Y = yaw
gender classification
result
test images
LBP feature extraction
SVM
online
45
testing images
0
+45
+90
Fig. 4 Samples of face images with seven facial expressions at ve face poses
from Radboud face database
45
Pose (deg)
0
+45 +90
Mean
95.3
92.9
90.4
88.1
97.4
94.1
92.4
90.3
85.3
Proposed method
IDP + proposed method
Gabor + proposed method
LPQ + proposed method
90
96.1
94.8
91.1
Pose (deg)
45
0
+45
99.1 100 99.4
96.9 100 95.9
94.6 98.8 95
+90 Mean
96.9 98.3
93.6 96.2
89.7 93.8
92.4
84.6