Research On Face Recognition Methods Based On Integrated Learning

As an important part of biometric identification, face recognition has been a hottopic in the field of pattern recognition. The traditional classification methods arealways based on single classifier, it is very difficult to find the signle classifier withhigh recognition rate when face image affected by illumination, facial expression andgesture. But integrated learning theory shows that if the weak classifier slightly betterthan random guess can be found, the strong classifier with any precision can beconstructed. Therefor this paper studied new face rcognition methods based onintegrated learning to further improve recognition performance.We focus on the face recognition methods based on integrated learning on thefoundation of making a deep study of ensemble classification and face recognitionmethods. Research mainly involves face image preprocessing, feature extraction andclassification recognition, the main results are as follows:(1) We propose a new face recognition method based on image contour. Firstly,facial contour energy image is obtained, and then we define a new similarity standardabout images, finally, classify the face image using this similarity. Comparativeexperiment with classical methods shows that our method has better performance andit is robust to the change of illumination and facial expression in some extent.(2) This paper proposes a new integrated face recognition method based onwavelet transform and two directional two dimensional principal component analysis((2D)2PCA). Firstly, this method extracts face features using discrete wavelettransform, then reduces dimensionality further with (2D)2PCA method. Finally, weclassify the test samples with boosting algorithm. Since boosting is an integratedapproach based on adjusting sample weights, complementary classifiers can beobtained further, it can improve the recognition accuracy when integrate theseclassifiers by voting method.(3) An integrated face recognition method based local features is proposed. Wedivide the face image into sub-image with same size in accordance with the theory offace segmentation, select the features as the attribute set in Attribute Bagging (AB)method, then train the base classifier based on different attribute sets, and finallyclassify the test sample with integrted classifier. The distribution of samples classifiedcorrectly by base classifiers is relatively uniform, this method can improve therecognition performance significantly by voting method. In the existing theoretical basis of integerated learning, we derive the training error boundary of multi-classAttribute Bagging method.(4) We propose an integrated face recognition method based on feature fusion.Firstly, this method obtains the global features using principal component analysis,sparsity preserving projections and simple projections mehods respectively, and getslocal feature information with sub-image method, these features are selected asattribute set to train the base classifiers, finally combines the base classifiers by votingmethod to classify the test sample. Because the method can simultaneously utilize theadvantages of global features and local features, and the Attribute Bagging methodcan significantly improve the classification performance, the new method has betterrecognition performance by comparing with global feature methods and local featuremethod.The experiments are conductued on Yale, Yale B and AR face databases to verifythe feasibility and validity of above face methods, these methods all achieve betterrecognition performance and experiment results are also analyzed.