Face Recognition With Multiple Variations Using Deep Learning

Shallow architectures are applied in a lot of existing machine learning algorithms, including neural networks with one hidden layer, kernel regression, support vector machine, and so on. Psychology research shows that, under the limited sample and calculating unit, these shallow structures are unable to express more complex function, especially to deal with the issue of classification problem with the visual input signals. Deep learning of nonlinear network structure, achieves complex function approximation, realizes distributed representation of the input data, while reflecting on the extraction capabilities from essential characteristics of the input sample data. By the influence of pose variations, expression and low resolution, which resulting in the rapid decline of the face recognition performance, pose variation particularly brings significant challenges to face recognition. Therefore, our proposed deep learning method using multi-layers nonlinear neural network, succeed to deal with face recognition under single or multiple changes from expression, pose and resolution.In this paper, deep network is used to map the original high-dimensional data into a low-dimensional feature data, in order to gather the images with different attributes belong to the same person as closely together as possible. Two neural networks with the same architecture, map frontal and profile face images to low-dimensional feature space respectively, where keep the local neighborhood information, to achieve multiple variations face recognition method. Experiments verify our proposed nonlinear locality preserving method can address the problem of non-frontal face recognition with expression variation under low resolution. Firstly, we perform the experiments on non-face database to verify the capability of preserving the neighborhood structure between the image data points in the original and feature space. The next experiments on several face databases show that superior recognition performance of our approach over the latest linear (or locally linear) methods.This paper is a study about multiple variation face recognition based on deep learning. In addition to the proposed approach mentioned above, we also seek some potential applications, and presence some value thinking which are promising in the future research.