Face Recognition Method Based On Statistics

Statistical pattern recognition methods have been successfully aPplied to many objectrecognition problems. One tyPical examPle is face recognition, which is the one of themost important fields in patt6rn recognition. The most conventional statistical-basedmethods include the eigenface method proposcd by Turk and PentIand and fisherfacemethod proposed by Be1humeur. In this thesis, we propose the eigenmotion method andnormalized LDA method.The first part of this thesis includes the first three chapters. It introduces the history of facerecognition research, especially the statistical-based methods such as the eigenface methodand the fisherface method.The second part of this thesis includes tbe ChaPter 4 and ChaPter 5. As we know, thedifficulty of the face recognition problem is to handle different types of variations, such asfacial expression, illumination and pose. In order to imProve the robusiness of facerecognition with respect to facial expression, this thesis proposes a new aPproach, theeigenmotion-based method, which is tOlerant to large variations of facial expressions. Thisnew aPproach first comPutes the motion vectors between a tCst face image and a neutraltraining image using the block-matching metbod, tben projects the motion vectors to a lowdimensional subspace that is pre-trained by aPplying Principal Component Analysis (PCA)to motion vectors resulting from training imagcs with expression variations. This subspaceis called an eigenmotion space. Finaly the identification of the test image is determined byits residue to the eigenmotion space. Both the individual modeling method and the commonmodeling method are described. At the same time, we consider the reconstructed ermrs of atest sample in tWo spaccs: the betWeen-class eigenmotion subspace and the within-classeigenmotion subspace, which are used as the classifier rule, in contrast to the traditionalmethods such as Euclidean distance or Mahalanobis distance in one subspace.Experimental results show that this method outPerfOrms the eigenface method in thepresence of facial expression variations. The aPproach can be extended to model othertypes of variations as well, fOr examPle, illumination and pose variations.The third part includes chaPter six, which introduces the normalized-LDA method. ltovercomes the drawback existed in the traditional LDA method. It redefines thebetween-class scatter by adding a weight function according to the between-class distance.TherefOre, it can separate the classes as much as possible. At the same time, it prOjects thebetween-class scatter into the null space of the within-class scatter that contains the most discriminant information. Hence, the transformation matrix composed with the eigenvectors corresponding to the largest eigenvalues of the transferred between-class scatter can maximize the Fisher Criteria. Experimental results show this method achieves better performance in face recognition than the traditional LDA method. It can also be applied on other problems of image recognition.The last chapter is the summary of this thesis. Also we present the difficulties in face recognition problem and the future work.