Research On Fast Face Recognition Based On Manifold Learning

With the rapid development of computer technology and the growing safetyneeds of the public, identity recognition has been attached a great of attention to, andface recognition, as a typical way of biometric feature recognition, is becoming one ofthe research hotspots. Although achievements have been made in the area of facerecognition through the linear subspace method, studies show that the face image islikely to be distributed in a low-dimensional nonlinear submanifold embedded intohigh-dimensional image space, and face recognition is often regarded as the problemof high dimensional and small sample, so that subspace classifier got through thetraining of small sample set does not work very well. Based on manifold learning, thisthesis took a deep study on the dimensionality reduction method based onnon-parametric kernel manifold learning and the classification methods based onsemi-supervised manifold regularization.Main reseaches are as follows:(1) To research the dimensionality reduction method based on non-parametrickernel learning and spectral regressionThrough researching manifold learning theories and techniques in featureextraction, and considering that the selection and construction of kernel function inthe graph embedding and kernel extension method affect the method’s performance,and that manifold learning and dimensionality reduction usually involves eigendecomposition probelm of dense matrix which is with high complexity in computingtime and space, this thesis, based on the dimensionality reduction framework ofmanifold learning and graph embedding, introducing non-parametric kernel learningand spectral regression method, proposes dimensionality reduction method based onnon-parametric kernel learning and spectral regression, which is highly efficient andextensible since it effectively avoids the problem of kernel fuction’s selection andconstruction for specific issue.(2) To research extrem learning machine method based on semi-supervisedmanifold regularizationFor the high computational complexity of kernel-based semi-supervised learningmethod, combining manifold regularization and pairwise constraint information, aswell as fast-learning-ability extrem learning machine method, this thesis proposes anovel extrem learning machine method based on semi-supervised manifoldregularization-a similar one to traditional semi-supervised method-which is not only applicable in semi-supervised conditionan, but also to large-scale learning tasks for itsdecision function, the experimental results on the real datasets demonstrate theeffectiveness of the proposed method.(3) To Design and implemente face recognition prototype system based onmanifold learningBy improving the dimensionality reduction method of face data and theclassification method of face recognition during the process of face recognition, thisthesis designs and implementes the prototype system which is based on bothnon-parametric kernel learning and the extrem learning machine based onsemi-supervised manifold regularization, and completes the software system designthrough using ORL, Yale and face database as the experimental data.