Study On Subspace Based Face Recognition Algorithms

With the development of information technology, one problem is inevitablyconfronted with large volumes of high-dimensional data when doing research inmodern science. The inner structure of the data is usually complicated, so that peoplecan hardly understand it by direct-viewing cognition. And then how to effectivelyextract the useful information becomes the key to solve this problem. With theproceeding of research, a number of effective face recognition methods have beenproposed. Among of them, the subspace analysis-based algorithms are aroused wideconcerns due to theirs favorable properties, such as convenient computation andeffectiveness for identification. Now the subspace analysis-based algorithm hasbecome one of the main methods of facial feature extraction and face recognition.The traditional subspace methods such as PCA and LDA mainly consider globallinear characteristics of data set and can not effectively find out the nonlinearstructural characteristics of the data. Manifold learning can explore the inherentstructure of nonlinear distributed data by its nonlinear characteristics. However, whensolve the tasks of face recognition, the manifold learning methods generally showmany shortcomings, such as small sample size (SSS) and out-of-sample. In order toovercome these problems, study on subspace based face recognition algorithms hasbeen done in this thesis. Moreover, the design and realize of face recognition based onARM embedded system has been built. The main works of the dissertation include:1. Isometric projection is a linear approximation to ISOMAP, which has localitypreserving properties, but it doesn’t consider the class discriminant information ofsamples. To overcome this problem, a novel subspace algorithm called discriminantIsometric projection (DIsoP) is proposed. Based on IsoP, maximum scatter differencecriterion (MSDC) is introduced to its objective function. After being embedded into alow-dimensional subspace, the samples of the different classes are far from each other.And then extracts the most discriminative feature. In addition, the algorithm avoids the small sample problem. Experimental results on ORL and Yale database demonstratethe effectiveness of the proposed DIsoP algorithm.2. The basis vectors obtained by IsoP are statistically correlated, and so theextracted features contain redundancy, which may distort the distribution of thefeatures and even dramatically degrade recognition performance. Due to this problem,a novel feature extraction method called Uncorrelated Discriminant IsometricProjections (UDIsoP) is proposed. Based on IsoP, UDIsoP takes the class labelinformation into account by constructing discriminant weighting matrix. At the sametime, the method obtains statistically uncorrelated features with minimum redundancyby introducing a simple uncorrelated constraint on the computation of the basis vectors,which makes it achieve better separability. Experimental results demonstrate theeffectiveness of the proposed UDIsoP algorithm.3. The dissertation takes ARM as the embedded system development plateform torealize a simple. And give the process of design and implementation to the embeddedface recognition system. Firstly the system completes the training of face recognitionclassifier and then the trained classifier results are transmitted to the above embeddedplatform. At last, realize the registration and certification based on the WinCEoperating system. Thus, we can conclude that the system has good real-timeperformance and accuracy.