Local Facial Action Units Recognition Based On DWT-KPCA

With the rapid development of the Internet technology and the maturity of network security, people look forward to a more harmonious human-computer interaction environment in which the computer can make corresponding reaction to the human expression. This requirement above promoted the rapid development of facial Action Unit (AU). AU reflects the more subtle levels of facial expression characteristics, and different combinations of AU not only can represent the six basic facial expressions, but also can better reflect the expression of the natural life. Thus, AU recognition had become one of the current research focus.Through the study of the extensive literature, we found that in order to study the AU, the face can be divided into upper and lower part. Then, the typical AAM algorithm was adopted to extract the AU deformation characteristics. Last, the classic SVM algorithm was used to classify AU. However, AAM algorithm was needed to mark manually, which cause heavy workload and limit the real-time performance. At the same time, there were many kinds of basic and combination AU which makes the results of AU recognition unsatisfactory. SVM algorithm has two important parameters to be analyzed and adjusted, which makes the AU classifying more difficult.In order to solve the problems above, this paper proposed to segment the face into sub-regions, in which the corresponding AU was studied respectively according to the corresponding relations of AU with a specific facial muscle movement. We used the Haar-like features to get the location of the eyes which was chosen as the basic to discribe the geometry relationship and segment the sub-regions which included eyebrows and eyes region、nose region and mouth region. After that, the adaptive threshold technology and projection of local eyebrow and eye region were introduced to find the proper bundary between the eyebrows and eyes.In the stage of feature extraction, we studied and alnalyzed the limitation of the classic PCA algorithm firstly, then proposed to use the KPCA algorithm. DWT (discrete wavelet change) was used to decompose the image, concentrate the energy, reduce computation cost and improve the calculation speed. Then, we put forward the improved algorithm that DWT-PCA and DWT-KPCA for feature extraction. According to analyzing the samples, we decided to adopt the K-nearest neighbor classifier for AU recognition. When the DWT-KPCA algorithm was used for feature extraction and the nearest neighbor algorithm was used for classifying, we got the best performance of AU recognition. The best average recognition rate was98.17%and the shortest average running time was0.05s for each sub-region image. The experiment results show that the algorithms we proposed achieved a satisfactory performance.