Face Recognition Algorithm Based On Singular Value Decomposition And Feature Fusion

Face recognition is an important part of the field of biometrics and a hot field of pattern recognition and machine learning,it has great academic and practical value.With the development of society and improving of people's living standards,applications of face recognition have become wider and wider,at the same time they put forward higher requirements.Face recognition is a complex and challenging research topic.Seeking stable,accurate,rapid face recognition algorithm is the development direction of the topic.Feature extraction is a key step in face recognition algorithm,which is also a study difficulty.This thesis mainly studies feature extraction.Firstly,Singular Value Decomposition(SVD)and its improved algorithms are studied,and then the thesis studies the global features and local features fusion.SVD feature describes the algebraic features of images.It has a good stability and rotation,displacement,shift invariance.But we find that SVD brings low success rate when it is used directly,on ORL face database recognition rate is only about 50%.Based on the SVD analysis,we find image subspace can improve sample's distance between the class,and SVD algorithm is improved based on sub-image,wavelet subspace and a combination of both.Based on improved SVD algorithm,we put forward an adaptive weights feature fusion algorithm.In general,the main contributions of this thesis can be summarized as follows.(1)Sub-image method and SVD algorithm combining were studied,unlike direct use of singular value,face image was firstly divided into a plurality of equal-sized non-overlapping sub-images.The thesis extracted singular value of each sub-image as a linear combination for face image recognition feature,and experimentally analyzed the number of sub-image and the number of singular value.(2)Wavelet subspace and SVD combining were studied,the image was divided into different sub-space by wavelet transform,sub-space singular value was extracted as a linear combination of face image recognition feature,wavelet function and wavelet layers were analyzed.(3)Based on the above two improved algorithms,this thesis proposed a new SVD algorithm combining spatial and frequency domain.Firstly,the divided sub-image was transformed by multi-layer wavelet transform.Every wavelet component is extracted by SVD feature,and the thesis chooses the largest linear singular value combination as the feature of identification and classification.(4)A feature fusion algorithm based on adaptive weights distribution was proposed.The algorithm used the improved SVD to extract global features and six key areas local feature.Based on adaptive weights distribution,six key areas local feature were combined into a local feature.Global and local features fusion was achieved in the feature integration layer.(5)BP neural network that based on resilient backpropagation(Rprop)algorithm is used to implement classification.Compared with the traditional algorithm,the Rprop algorithm has fewer training times and high operation efficiency.