A Study Of Face Recognition Approach Based On Semi-supervised Discriminant Analysis

Recently, face recognition(FR) technology has achieved encouraging results.However, FR is still a challenging research topic because of the complicated facial variations.The key problem of FR is the extraction of robust facial features. FR methods can be mainly divided into two categories in terms of feature extraction, namely the global feature extraction and local feature extraction. The representative methods for global feature extraction contain Principle Component Analysis(PCA) and Linear Discriminant Analysis(LDA), while the typical algorithms for local feature extraction are Locality Preserving Projection(LPP) and Unsupervised Discriminant Projection(UDP). Global feature and local feature have its own advantages in face recognition. Global feature can deal well with facial images under uniform illumination changes, but it is sensitive to variations in pose and facial expression. Local feature is insensitive to local appearance changes caused by facial expression and pose, but not robust for uniform illumination changes. However, most of the face recognition algorithms did not consider these two kinds of features simultaneously. This means that they are not able to make full use of the complementary advantages of these two facial features,and thus their performance will be negatively affected.This dissertation is comprised of four chapters. The main contributions of the work are included in chapter 2 and chapter 3. Chapter 2 proposes a semi-supervised discriminant analysis(SSDA) FR method based on the feature non-linear fusion strategy. We first establish a discriminant criterion by non-linearly combining the global feature and local feature, and then develop a cross iterative algorithm via maximizing the discriminant function, which is theoretically shown to be convergence. The optimal parameters and projection matrix can be determined automatically at the same time using the proposed cross iterative algorithm.Moreover, the combination parameters are guaranteed to fall into the interval ]1,0[.The proposed method not only integrate the complementary advantages of the two kinds of facial features, but also further enhance the performance of SSDA by using the geometric distribution weight information of the training data.Three publicly available face databases,namely ORL,FERET and CMU PIE face database, are selected for evaluations. Compared with some state-of-the-art FR algorithms, the experimental results demonstrate that theproposed SSDA method gives superior performance.To solve the 3S problem of SSDA approach, chapter 3 proposes a novel semi-supervised regularization method(RSSDA). It is theoretically shown that the regularization matrix converges to the original matrix in Frobenius norm as the regularization parameter tends to zero. This method is actually conducted in the entire feature space and able to avoid losing useful discriminative information. So, we can obtain the optimal solution in the whole pattern space. The proposed RSSDA method is tested on two face databases, namely ORL and FERET databases. Compared with some LDA-based and LPP-based approaches,experimental results show that our method outperforms LDA, LPP, UDP, SSLDA, FFS and SSDA approaches.