Feature Extraction Based on LSDA for Lipreading

This paper proposed a new feature extraction method for lip-reading, named DCT+LSDA. Discrete Cosine Transform (DCT) is a popular method used to reduce the dimension of the data and it has been very efficient in lipreading. Linear Discriminant Analysis (LDA) is a method to study the class relationship between data points, it is very useful method for dimensionality reduction and feature extraction. For lipreading system, the change of the lip is a non-rigid deformation, only considering the discrimination of different class is not enough, the local structure information is important too. So in this paper, the Locality Sensitive Discriminate analysis (LSDA) is used, it is a method considers both the Discriminant and geometrical structure of the data. The experimental results show that the proposed method DCT+LSDA is performed better than DCT+PCA and DCT+LDA, and it also shows that the endpoint detection is crucial for the lipreading system.