Research On Face Recognition With Single Training Image Per Person

As one of the most successful applications of image processing and analysis in the field of biometric authentication, Automatic Face Recognition (AFR) is now a hotspot and attracting more and more scholars to research it. AFR is more convenient and friendly than other biological authentication technology. But in many application conditions there is only one training image per person, in this case most of the reported face recognition algorithms will suffer serious performance drop or even fail to work. To do special research on face recognition with single training image will be the profound significance of the AFR application and theoretical research.This thesis firstly introduced the background and significance of the face recognition research subject, summarized and analyzed the status quo and progress of the studies on this subject at home and abroad. A new face recognition approach based virtual samples generation and face characteristics enhancement was proposed after the detailed analysis of the specificity of the single training samples face recognition problem. Candide-3 model was applied to the novel method above to generate numbers of virtual samples with different poses by single image. This new algorithm also combined with the method of face characteristics representation and enhancement which base on the singular value decomposition (SVD) under a unity fundamental space, made use of a simple nearest neighbor classification method to classify the faces. The new method achieved good results, and eased the problem of serious performance drop of face recognition rate under the single sample conditions.In addition, the AFR technology includes two tasks: face identification and face verification. However, most of the studies focus on face identification. Only a few researchers study the difference between face verification and identification deeply. Based on this reality, this thesis focused on the automatic face verification technology's specificity, and then proposed a useful method of face verification with single training sample face, which based on Support Vector Machines (SVM) and Candide-3. The new method firstly got a human face's Candide-3 model and then reconstructed different figures of that face by the model. After that trained all original samples and new reconstructed samples together and classified them by SVM. Experiments on the ORL face base showed that this new method obtained higher recognition performance than the regular SVM method.The forth contribution of this thesis is the design of SVM-based verification system to meet the practical need of the verification of candidates who take part in an examination. This thesis thoroughly discussed the common thing of face verification system such as camera environment, key recognition algorithms, threshold setting and so on.