Facial Expression Recognition with Discriminative Common Vector
 
 
Chun-Hao Huang (黃俊豪)
中華民國 96 年 7 月
輔仁大學---智慧型系統實驗室---雲端視覺組
 
 
 
 

Overview

  Recently facial expression recognition system has become an important issue in both human-computer interaction (HCI) and human-robot interaction (HRI). It is an important issue to extract features from face images to recognize facial expression. In this paper, we apply a face feature extraction approach, namely discriminative common vectors, for the recognition of the six basic expressions including happy, sad, angry, disgust, fear and surprise. By applying discriminative common vector, we can reduce the dimensionality of image and classify them in a lower dimension which would be useful in later recognition procedure. Then we use HMM as our classifier to find the time series information of the feature vector projected by discriminative  common vector.

 

Method

   We use spatio-temporal information for recognition of expression, which considered image sequence as an expression. Therefore, we utilize optical flow images as our primitive features which is computed from two images and contain the time-series information. In order to reduce dimension, an discriminative common vector (DCV) method is adopted to classify images in spatial domain. We then extend the methods to apply it in spatio-temporal domain by dividing a image sequence into several parts. Each part of the sequence denotes a level of expression, namely facial states. Facial states are classified by Temporal-DCV and can be considered as hidden states. Therefore, we use Hidden Markov Model (HMM) as our classifier for recognition of time-series information and facial states can be considered as the hidden states in HMM. We test number of facial states for getting optimal state number. And a comparison of classifying by HMM and voting scheme is also compared to show the performance of the proposed method.

 
 

 

 

 

 

 

 

Conclusions

   In this work, we use discriminative common vector to extract feature vector from video sequence. The use of the DCV can not only reduce dimensionality of input image but also cluster them. It is shown that dimension reduction by Temporal-DCV can have better accuracy to extract time series information when coupled with HMM. And DCV can overcome the difficulty of the situation when dimensionality of image is larger than the sample of images. It is desirable that a functionality to segment image sequence in a continuous expression image automatically be implemented to have a better use of the system.

 
 
 
輔仁大學---智慧型系統實驗室