Face recognition using AdaBoost modular locality preserving projections

Locality Preserving Projections (LPP) is a linear manifold learning method proposed for feature extraction which can optimally preserve the neighborhood structure of a data set. Although LPP has been widely applied in image recognition, working as a holistic approach, it is sensitive to variations caused by illumination and expression, and to inaccuracy in face localization. To alleviate these problems, in this paper, we propose to combine the block-based LPP with the AdaBoost algorithm to select the features to improve the accuracy of face recognition. This algorithm, namely AdaBoost Modular LPP (AMLPP), divides a facial image into many overlapping small blocks, and applies LPP to these blocks to form block features. By using'pseudo-loss' and by updating the distribution of mislabelled samples in the AdaBoost algorithm, AMLPP selects adaptively those optimal block features from a huge set of potential block features to form a number of weak classifiers, which are then combined for the construction of a strong classifier for accurate and efficient face recognition. In each feature-selection process, optimal features are selected to generate weak classifiers, which emphasizes those hard-to-classify samples. Our AMLPP algorithm is compared with the LPP algorithm, the neighborhood preserving embedding (NPE), the discriminant locality preserving projections (DLPP), and the orthogonal locality preserving projections (OLPP), based on the Yale and YaleB face databases. Experimental results show a significant improvement when using our proposed algorithm.