A Study Of Face Recognition With Single Sample Per Person Based On Variation Dictionary

With the rapid development of artificial intelligence, face recognition(FR), especially the technologies of FR based on multiple training samples per person, have been applied in practical application with significant performance improvement. However, in some applications, such as passport verification, law enforcement and identification verification, each person can only capture a single training sample without the important within-class variance, which makes FR algorithms based on multiple training samples unavailable. Thus FR with single sampler per person(FRSSPP) has been an urgent and challenging research topic.By estimating face variations from a collected generic training set, the best performance has been achieved in FRSSPP. However, the accuracy of FRSSPP can still not meet the demand of practical application, especially the recognition rate of FRSSPP with the variations of expression, pose, occlusion and misalignment is still low. Based on the generic training set, this thesis proposes algorithms of FRSSPP based on local facial feature and variance dictionary learning, achieving visibly better performance than prevailing methods. Our contributions are as follows.Firstly,we propose a local variation joint representation(LVJR) method, which learns a variation dictionary and does joint and local collaborative representation for a query image. The learned variation dictionary is required to do similar representation for the same-type facial variations, while joint and local collaborative representation can effectively use local information of face images. Experiments on the large-scale CMU Multi-PIE and AR databases demonstrate that the proposed LVJR method achieves better results compared to the existing solutions to the single sample per person problem.Secondly, we propose a triple local feature based collaborative representation(TLC) method to make full use of the local information. We firstly extract different types of Gabor features for the face image by using Gabor filters with different scales and different directions. Then we partition each type of Gabor feature into several local patches to obtain the triple local features, which intrinsically encode the local information of local scale, local orientation and local space. Finally, we do the local collaborative representation and the classification based on these triple local features. Experiments on the CMU Multi-PIE and AR databases demonstrate the superiority of TLC over the current state-of-the-art methods.In conclusion, this thesis proposes a variance-dictionary FR with single sample per person. Through our depth research on FRSSPP,we find that discriminative local features and variance dictionary learning can significantly improve the recognition accuracy of FRSSPP, which is a very promising research direction.