Recognizing Facial Expression Based On The Fusion Of Multi-order Fractional Fourier Domain Features

Recognizing facial expression by computer has been an attractive issue in areas such as image processing, computer vision and pattern recognition, which brings people practical and economic value in the fields of video conference, artificial intelligence and digital family etc. In this paper, it mainly makes researches on facial expression recognition technology based on the fusion of multi-order fractional Fourier domain features. Aiming at solving the disadvantages of traditional fusion methods, it proposes a novel feature fusion algorithm-Discriminative Multi-set Canonical Correlation Analysis. Then it is used to fuse multi-order fractional Fourier domain features of facial expression images. Recognition accuracy is greatly increased, meanwhile the dimensions of fused features are dramatically reduced, and so real-time processing is guaranteed. The main work of this paper is organized as follows:1. It describes the Fractional Fourier Transform (FrFT) and the discrete algorithms and two dimensions form of FrFT. Moreover, facial expression recognition based on single-order fractional Fourier domain features is discussed. It concludes that single-order fractional Fourier domain features are always not enough for recognition, so feature fusion is necessary.2. It proposes a novel feature fusion algorithm-DMCCA. As a widely used fusion algorithm, Canonical Correlation Analysis (CCA) can extract and fuse associated information between two sets. However, CCA ignores discriminative information among classes which can benefit classification. What’s more, CCA only can fuse two sets and fails to fuse more. Aiming at solving the two disadvantages of CCA, this paper proposes a novel feature fusion algorithm-DMCCA. It can not only extract and fuse discriminative information which can benefit classification, but also fuse more than two sets, which makes it more pervasive. Moreover, it proposes that the K-L transform can solve the problem of high dimension and small sample situation, which clear obstacles for widespread use of DMCCA.3. It proposes a facial expression recognition method based on fusion of multi-order fractional Fourier domain features. The serial fusion, CCA and DMCCA are employed to fuse multi-order fractional Fourier domain features of facial expression images. Simulation results show that no matter which method can increase the recognition rate relative to that of single-order when fuse certain features. It explains that the fusion of multi-order fractional Fourier domain features increase the amount of effective information which is necessary for classification and thus the recognition accuracy is more exact. In addition, among the three fusion algorithm, the fusion effect of DMCCA is the best. It can extract and fuse discriminative information which can benefit classification, and reduce the feature dimension and guarantee the real-time processing at the same time. This paper also discusses the facial expression recognition based on fusion of different orientations and scales of Gabor features and compares it with that based on multi-order of FrFT. Simulation results show the effectiveness of DMCCA and the superiority of fusion of multi-order of FrFT features.