Information Technology Reference
In-Depth Information
region of the input space and multiple classifiers can supplement each other. According
to our best knowledge, only few studies in facial expression recognition paid attention
to multi-classifier fusion. To utilize the advantage of the multi-classifier fusion, in this
paper, we also extend a framework of multi-classifier fusion based on decision rules to
facial expression recognition.
In this paper, we propose a novel component-based approach for facial expression
recognition from video sequences. Inspired by the methods presented in [15,21], 38 im-
portant facial interest regions based on prior information are first determined, and then
spatiotemporal feature descriptors are used to describe facial expressions from these ar-
eas. Furthermore, we use AdaBoost to select the most important discriminative features
for all components. In the classification step, we present a framework for fusing recog-
nition results from several classifiers, such as support vector machines, boosting, Fisher
discriminant classifier for exploiting the complementary information among different
classifiers. Extensive experiments on the Cohn-Kanade facial expression database [22]
are carried out to evaluate the performance of the proposed approach.
2
Boosted Component-Based Spatiotemporal Feature Descriptor
2.1
Facial Interest Points
In many earlier methods [1,2,23], fusion of geometric features and appearance features
can improve the performance of expression recognizers. Geometric features are usually
formed by parameters obtained by tracking facial action units or facial points' variation.
It is well known that not all features from the whole face are critical to expression
recognizers. Yesin et al. [14] proposed to apply optical flow to regions based on posi-
tions of the eyes, eyebrows and the mouth. Zhang et al. [24] developed a framework
in which Gabor wavelet coefficients were extracted from 34 fiducial points in the face
image. In methods on scale-invariant feature transform (SIFT) [21], SIFT keypoints of
objects are first extracted from a set of reference images in order to avoid from comput-
ing all points in an image. It is thus found that the search of interest points or regions in
facial images is more important to component-based approach.
However, faces are different from other objects, in other words, important features
for facial expression are always expressed in some special regions, such as mouth, cheek
etc. Thus different from SIFT, our interest points detection is based on prior-experience.
In our paper, 38 facial points are considered, shown in Fig. 1(a).
The approach for detecting those interest points is critical to our approach. If con-
sidering accuracy, manual labeling facial points for face image is good for expression
recognizers. Unfortunately, this method costs much time and is not practical. It is well
known that some methods are proposed for detecting or tracking facial points, such as
AAM, ASM, and Elastic Bunch Graph Matching etc. After comparison, ASM [25] is
applied to detect the facial points.
Geometric information from the first frame is obtained by applying ASM as shown
in Fig. 1(a). Here, geometric models are trained from FRAV2D database [26] and MMI
database [27].
Search WWH ::
Custom Search