The research on Automatic Face Recognition (AFR) has both significant academic importance and wide applications. Based on statistical learning and fusion, the three key problems in AFR--- alignment, feature representation and training and construction of classifier, are studied in this thesis deeply. With consideration on particularity of human face image and small sample size problem in statistical face recognition, this dissertation incorporates the technology of fusion into statistical learning so as to improve performance of AFR. And the main contribution of this thesis includes: Firstly, a multiple-classifier based eye location method is proposed from the viewpoint of fusion. Totally four eye detectors with different properties are trained by statistical learning algorithm Boosting. Then Dempster-shafer theory is used to combine outputs of the four eye detectors and decide true eye location. Experimental results demonstrate this eye location method is precise. Secondly, a Weighted Gabor Complex Feature (WGCF) is proposed to fully make use of discriminant information extracted by Gabor wavelet. The Gabor magnitude and phase feature vector are combined into one complex vector by proper weights so as to use both kinds of information. Meanwhile, the subspace based recognition algorithms, Principle Component Analysis (PCA) and Linear Discriminant Analysis (LDA) in Euclidean space, are generalized into Unitary space, which makes WGCF easily be used with them. Experimental results show that WGCF is better than Gabor magnitude and phase feature. Thirdly, with consideration on speciality of human face pattern and robustness of feature representation, a face recognition algorithm based on Dissociated Region Pair Linear Discriminant Analysis (DRP-LDA) feature is proposed. Multiple dissociated region pairs (DRP) are used to represent a face image and then the Linear Discriminant Analysis (LDA) feature is extracted from each DRP. Similarity of two faces is the weighted sum of similarities of LDA features of multiple DRPs that are used to represent a face.The parameters including DRP positions and weight coefficients are learned by data driven method in the framework of statistical learning. Experimental results show the algorithm is effective and the DRP representation is powerful. Fourthly, a Boosting in random subspaces face recognition method is proposed. Multiple random subspaces are generated from original feature space and training is conducted in each feature subspace independently. Then strong classifiers trained from multiple random subspaces are combined into one stronger classifier for face recognition. This method improves efficiency of Boosting, especially training process, greatly. Extensive experimental results show that the method is reasonable and effective.
修改评论