Research And Application For Face Recognition Based On Ensemble Learning

Face Recognition as an important part of pattern recognition research, it is safe, direct and concealed when compare with other identification technology. Recently, the researches on face recognition are focus on which features can be used to represent a face and how to classify a new face images based on the chosen representation.This dissertation introduces the current research situation and main methods for face recognition and studies the key technology for face recognition, then proposes Regularized Linear Discriminat Analysis based on ensemble learning and Regularized Compound Kernel Linear Discriminat Analysis based on ensemble learning for face recognition.First, a regularized Fisher's discriminate criterion is used to address the Small Sample Size (SSS) problem. Thought adjust the regularization parameter to decreasing the larger Eigenvalues and increasing the smaller ones, thereby counteracting the biasing and overcome the SSS problem.Second, based on the Linear Discriminate Analysis (LDA), the method of compound kernel function is proposed. By using kernel function, the transform of low-dimension nonlinear space to high-dimension linear space is realized by nonlinear mapping and the face features are extracted in the high-dimension space by using approved linear discriminate analysis to form nonlinear optimal features.Last, the ensemble learning algorithms combine the weak classifiers into composite classifiers, which is more robust than the original one. In the dissertation, Adaboost algorithm is used. Its multi-class extensions include two variants, adaboost.ml and adaboost.m2. Adaboost.ml is most straightforward generalization, however,the algorithm halts if the classification error rate of the weak classifier produced in any iterative step is>50%. In order to avoid this problem, adaboost.m2 brings in pseudo loss. With the pseudo loss the Boosting process can continue as long as the weak classifier produced has pseudo loss slightly better than random guessing. In this way, adaboost.m2 can focus the leaner not only on hard-to-classify examples, but more specifically, on the incorrect labels.The method based on ensemble learning is take advantage of both Boosting and LDA. Weighting functions is introduced into the samples which closer in the output space in each iteration, So that the class separability is enhanced in the new feature subspace.