Multi-view Manifold Discriminant Learning Based Single Sample Face Recognition

Recently, with the development of Internet and big data technology supported by the country, face recognition has drawn much attention from many fields, including intelligent security, criminal investigation, Internet payment, etc. Most face recognition methods perform very well with adequate training samples. However, in real-world applications each person sometimes has only one training sample, in which most face recognition methods decline sharply. Therefore, single sample face recognition(SSFR) is very important. To solve this problem, this thesis proposes several effective SSFR approaches from the perspective of multi-view manifold discriminant learning.Firstly, this thesis proposes the multiple manifold discriminant learning based SSFR approach(MMDL). This approach first extracts multiple facial variation features from the generic set, add these variation features into the original training set, and obtain the extended sample set. Then, each extended facial image is partitioned into several nonoverlapping patches, which can construct a manifold for one class. SSFR is transformed into the multi-manifold matching problem. MMDL can learn multiple discriminant feature projection matrixes, which make the manifold margin between inter-class maximized and the manifold difference of intra-class minimized, simultaneously.Secondly, to well exploit the multiple features from facial images, this thesis further extracts multi-view features from facial images and proposes the multi-view manifold discriminant learning based SSFR approach(MVMDL). It utilizes the similar manner to obtain the extended sample set as in MMDL. Then it extracts multi-view features, and builds the multi-view manifold discriminant learning model to learn multiple discriminant projection matrixes.Lastly, to make the training sample extension and the manifold matching more accurate, this thesis proposes the collaborative representation and multiple manifold learning based SSFR approach(CR-MVMDL). It utilizes the collaborative representation classification method to select the similar sample set from the generic set for each training sample. Then the training set extension and the manifold matching are conducted within the similar sample set. This approach can make the training sample extension more accurate and the projection matrixes more discriminant.Experiments on the AR, ORL and LFW datasets illustrate the effectiveness of proposed approaches. And they outperform all compared methods in terms of recognition accuracy.