Face Recognition Based On Graph Embedding

Face recognition has received increasing attentions in pattern recognition and machine learning due to its special advantages. However, the high-dimensionality of unstructured data is computationally challenging to analysis, and face image is also one kind of high-dimensionality data. To manage this problem, many dimension reduced methods have been proposed. They are the important part of face recognition technology and the key to solve the problem of image recognition. In this paper, our attentions focus on the algorithms based on graph embedding, then we analysis several classical graph embedding methods in detail, and two feature extraction algorithms based on graph embedding are proposed. The main work and innovations of this dissertation are summarized as follows:(1) Local discriminant embedding(LDE) attempts to achieve high recognition accuracy, implicitly assuming that all misclassifications lead to the same losses. This assumption, however, may not hold in the practical face recognition systems, because the losses of different mistakes may be different. Motivated by this concern, a new approach called cost-sensitive local discriminant embedding(CSLDE) is proposed in this paper. First the feature extraction phase utilizes the cost-sensitive learning technique which helps analysis different misclassifications by constructing the cost matrix. Then we maximize the costs of misclassifying the neighboring points of the different class and minimize the distances of neighboring points of the same class simultaneously. Finally we obtain the optimal orthogonal vectors which help maintain the metric structure by utilizing an iterative algorithm. The extensive experiments on the face database Yale, ORL, AR and Extended Yale B demonstrate the effectiveness of the proposed algorithm.(2) Feature extraction algorithm plays an important role in face recognition. However, the extracted features also have overlapping discriminant information. The property of statistical uncorrelated criterion eliminates the redundancy among the extracted discriminant features, while many algorithms generally ignore this property. In this paper, we introduce a novel feature extraction method called local uncorrelated local discriminant embedding(LULDE). The proposed approach can be seen as an extension of local discriminant embedding framework in three directions. First, a new local uncorrelated criterion is proposed, which effectively captures the local information of interclass and intraclass. Second, we reconstruct the affinity matrixes of intrinsic graph and penalty graph which mentioned in LDE to enhance the discriminant property. Finally, it overcomes the small-sample-size problem without utilizing principal component analysis to preprocess the original data, which avoids losing some discriminant information. Experimental results on Yale, ORL, Extended Yale B and FERET database demonstrate the performances of LULDE outperform LDE and other representative uncorrelated feature extraction methods.