Face And Expression Recognition Based On Ensemble Features

Face Recognition and Facial expression recognition has been widely used in many areas such as computer games and human-computer interaction but its analysis is still a challenging problem due to facial appearance variations in illumination, pose, ageing, partial occlusions and other changes.Deriving an effective facial representation from original face images is a vital step for successful facial expression recognition. Different features play different roles in extracting information and Face images contain too much information, so only a single model appears not abundant. In order to make full use of discriminative information in each feature model and improve the system performance, all of the models are combined to form an ensemble classifier of face recognition and expression recognition.This paper proposes a novel face recognition approach, in which firstly face images are represented by Gabor Pixel-Pattern-Based Texture Feature (GPPBTF) and Local Binary Pattern (LBP).Then Null Space-based Kernel Fisher Discriminant Analysis (NKFDA) is applied to the features of GPPBTF and LBP independently to obtain two result identifiers for each image. Eventually, the two result identifiers are combined to reach a final identification. Extensive experiments on FERET face databases demonstrate that the proposed method not only greatly reduces the dimensionality of face representation, but also achieves more robust result and higher recognition accuracy.We also propose a novel facial expression recognition framework by fusing five independent models in this paper. Firstly, face image is described by different feature descriptors:Local Binary Patterns (LBP), Histograms of Oriented Gradients (HOG), Gabor wavelets, and two pixel-pattern-based texture features (PPBTF). The aim of this work is to extract facial information as much as possible. Secondly, each descriptor is feeded into separate JointBoost classifier channel.Thirdly, classification results from each channel are integrated into a final result. The framework is evaluated on Cohn-Kanade database and compared with individual models. Experimental results demonstrate that the proposed method achieves considerable performance.