Font Size: a A A

Research On Face Recognition With Single Training Sample

Posted on:2013-12-27Degree:MasterType:Thesis
Country:ChinaCandidate:X Y XuFull Text:PDF
GTID:2248330362970887Subject:Computer software and theory
Abstract/Summary:PDF Full Text Request
Face recognition is an important subject in artificial intelligence field, and it has gained extensiveattention from researchers in the past decades. This paper focused on research of face recognition ofsingle sample, which starts from construction of virtual samples and work out different solutions inthe light of variable factors such as poses, illumination and facial expression. The creativity andresearch work of this paper are summarized as follows:(1) In view of face recognition of multiple poses, a method which can merge different picturesof poses of faces is put forward. DWT (Discrete Wavelet Transform) is adopted to removehigh–frequency information of caused by changes of facial expression and covers in pictures of faces.Experimental results on ORL face database demonstrate the effectiveness and advantages of theproposed method.(2) In the light of robustness of Gabor feature and LBP feature on changes of facial expressionand illumination, mutual information is introduced to measure the similarity of LGBP maps.Experiments show that the new method proposed gains different levels of improvement onperformance of recognition in changes of facial expression and illumination compared with othermethods such as PCA, LPS and LGBPHS.(3) The method of SIRS (Single Image Resampling Subspace) is proposed. A virtual image canbe constructed for every real image by interval sampling. Then subspace is constructed for everyimage and images are classified by the distance of subspaces. According to experiments, illuminationrobustness and facial expression robustness are regarded as good features of this method.(4) Based on traditional kNN (k-Nearest Neighbor) algorithm, FI-kNN (Fuzzy Inferencek-Nearest Neighbor) algorithm is proposed with introduction of fuzzy logic. To take into account theeffective nonlinear structure information, we further extend FI-kNN to its kernel version KFI-kNN.FI-KNN is applied to face recognition of single training sample and the effect of recognition isimproved to some extents.
Keywords/Search Tags:single training sample, face recognition, multiple poses, mutual information, Resampling, Fuzzy Inference k-Nearest Neighbor
PDF Full Text Request
Related items