Research Of Facial Landmark Location And Face Recognition

The object of face recognition (FR) is automatically validating and recognizinghuman identity with face image. FR is the most representative application in patternrecognition (PR), many pattern recognition methods supply FR with theoretical basis.At the same time, the development of FR also further widens the application area ofPR method, and speed up the development of PR. Some PR methods were proposedfor FR problems in the beginning, and then were used in other PR application areas.FR technique has a wide application in anti-terrorism, society safety, surveillance sys-tem, which makes it attract increasingly attention. However, there are many problemsneed to be solved before FR technique can be really applied in practice. Some ofthem are fairly apparent, such as facial landmark location, estimation of face pose, bal-ance between recognition rate and speed. Therefore, this thesis extensively studies theabove-mentioned problems, the main contributions contain:1. Detailing FR system developed by our research group and its sub-modules, whichinclude face detection, FR, facial landmark location, facial feature extraction,manifold learning, face pose estimation, 3D face model. Describing their func-tion in the whole system.2. Describing the development of FR, current research status, domestic and in-ternational research institutes, international FR contest and commercializationproducts, outstanding findings as well as mostly used and publicly available facedatabases.3. Deeply studying the facial landmark location methods ASM (active shape mod-els) and AAM (active appearance models). Analyzing the disadvantages of ASM.Proposing to extend the local profile in ASM from 1D to 2D such that it containsmore landmark information and improves the accuracy of location for each land-mark, and accordingly improve the whole accuracy of location. Proposing to convert the problem of finding new position from minimization of Mahalanobisdistance to a classification problem using SVM (support vector machine) clas-sifier. Proposing to use different similarity functions for different landmarksaccording to their specific feature. Proposing to constrain the displacement oflast level in the multi-resolution searching scheme, which decreases the effectof noise. Proposing to construct local profile with different lengths in differentlevels.4. Studying the representative subspace methods in FR:PCA(principal componentanalysis),LDA(linear discriminant analysis),ICA (independent componentanalysis) and LPP (locality preserving projection). Proposing to combine above-mentioned methods with AP(affinity propagation) to form new PR methods. Un-like tradition subspace methods which need to compare the test sample to eachtraining sample in the testing phase, the new PR methods only use representativesamples corresponding to each class for identification. Such kind of identifica-tion scheme not only avoids the effect of noise in the training set, but improvesthe recognition speed.5. Introducing some popular manifold leaning methods, and extensively studyingRBMNN(Restricted Boltzmann Machine Nerual Network) dimensionality reduc-tion method. Proposing to generate more samples by subsampling to deal withthe problem that RBMNN needs a great deal of data for training. Proposingto combine PCA with RBMNN to decrease the number of net nodes as well asthe training time. Proposing to using Gabor feature as input, use RBMNN fordimensionality reduction, and further improve the recognition performance.