Towards Face Recognition From One Single Training Sample Per Person Using Algebraic Features

Face recognition is a main branch of biometric recognition. And as it is non-contact and non-intrusive, face recognition has many obvious advantages over the other biometric technologies such as fingerprint recognition and iris recognition. Nowadays, with the agent security demands both domestic and abroad, face recognition is increasingly favoured by many application fields including public security and access control. The warming of the applications promotes the related studies, and face recognition study shows a brisk growing trend. However, some difficult problems of face recognition are still unsolved.This dissertation mainly studies the problem of face recognition with single sample. Single sample here is the short of"single training sample per person"which refers that: in the known face image database, such as the passport face image database, each person has only one face image of his own. It's the most extreme case of the small-sample-size problem, and also the most extreme case of the whole face recognition field. For face recognition with single training sample per person, many conventional face recognition methods which work with many training samples can't function well, and some methods can't work at all.The dissertation focuses on face recognition based on algebraic features with single sample per person. The main research work and contributions of it are as following:(1)On the basis of analyzing the storage format of grey images, a new sample augment method called bit-plane image sample augment method is proposed. In this method, each original grey image can be sliced into eight bit-plane images, among which the low-level sliced images are kicked out because of containing obvious noise, while the high-level sliced images have much discriminative information in them and are kept as the augmented samples. Then, each person's single original face image and the augmented face images from it together work as his new training samples. After augmentation, the discriminative information is successfully strengthened at some extent, which greatly benefits the follow-up processes: feature extraction and classification. Finally, experiments verify the effectiveness of the proposed algorithm.(2)According to the research thinking of transforming the single-sample face recognition problem into the common multi-sample face recognition problem, some analysis is made which is towards traditional sample augment methods and recognition algorithms. And find that some of the traditional algorithms can't function better mainly because the augmented samples contain too much noise. Then, aiming to eliminate the noise more effectively, a new sample augment method named generalized slide window (GSW) is proposed. And it is a linear discriminant analysis (LDA) based method. This method won't carry any noise into the new samples from the outside when it works, hence it improves the quality of the new samples to the best, which is greatly helpful for the follow-up feature extraction and recognition. Apart from that, the new sample augment method is more tractable than some other traditional ones. The effectiveness and advantages of the proposed method are demonstrated by experiments with the augmented samples.(3)With the consideration that there maybe exists outlier face images in the single-sample face image database, the pair-wise weighted idea is added to the former proposed frame of single-sample face recognition. The new frame has two advantages: firstly, the noise is effectively eliminated during the process of augmenting samples, thus better samples can be obtained. Secondly, the problem resulting from the outlier patterns is better solved in the process of feature extraction, which ensures that the features extracted are more discriminative. The former advantage realizes face recognition with single training sample per person. And the latter helps optimize the recognition performance: it is as good as the non-weighted method when there aren't outlier patterns, and much better than that in case of existing outlier patterns. So, on the whole, the performance of the new recognition frame becomes more effective and more robust. And experimental results on the ORL face image database confirm the above advantages.