Face Recognition Based On Linear Representation Model

Currently, the frequently-used face recognition algorithms are mostly relied on statistical learning theories. These algorithms achieve good effect in the processing of low dimensions. However, with the popularity of high-definition video of camera, more and more high dimensional image appear constantly. As for the traditional algorithms, it is difficult to overcome the defects in dealing with high-dimensional data, which thus leads to much bigger difficult to deal with the existing high dimensional image data.Compressed sensing theory provides a new idea to transform the big data from high dimension to low dimension and also provides a new direction to the developing of face recognition. The most representative work is the Sparse Representation-base Classification, SRC, proposed by Wright et al. SRC constructs a training dictionary with some known images of face, and testing images are represented as the linear combination of the dictionary, then reconstruct the testing images based on the coefficient of linear combination and classify testing images by error of reconstruction. Inspired by SRC, in this thesis, we study relevant issues of face recognition based linear representation model. The main research contents are summarized as follows:First of all, this thesis works on the dimension reduction problem of face images. The methods of dimension reduction can be divided into two broad categories:training and without-training. And then, we describe five commonly used dimension reduction methods in detail. We analyze the effect of those dimension reduction methods for face recognition through abundant experiments.Secondly, four important linear representation models based on the least square have been concretely studied. We concentrate on the solution procedure of every model, and finally apply them to face recognition. We compare the results on face recognition for every model and analyze the insufficient of linear representation models based on least square when handling the residual.Lastly, to the question of that commonly used linear representation of the linear representation model requiring maximum likelihood estimation on residuals, we study a representation model without residual estimation, called Dantzig Selector model, a very effective method for solving this model has been researched through detailed theoretical description and proof. Then, inspired by face recognition method based on sparse representation, we propose a face recognition algorithm based on Dantzig Selector.The experimental results adequately validate the effectiveness of proposed algorithm in aspect of face recognition.