Research On Face Recognition Based On Deep Learning

In practical applications, such as surveillance systems, collected face images captured by cameras are often low resolution and with great pose variations. Affected by pose variation and low resolution, performance of face recognition degrades sharply, while pose variations bring great challenge to face recognition. Pose variations bring nonlinear factors into face recognition. Architectures are applied in many existing machine learning algorithms including neural networks with only one hidden layer, kernel regression, support vector machine, and many others are using shallow architectures. Psychology results shows with limited samples and finite computing units, those shallow architectures are incapable of representing the complex function, and place restriction on the generalization capability of classifying complicated issues, especially for the rich sensory input. Deep learning achieves the approximation of complex function, characterization of the input data by learning a deep nonlinear network, and shows the great power in extracting the intrinsic feature of the training data. In this paper, we employs deep neural network to overcome the influence of pose variation and low resolution, present an approach to super resolution recognize the face with pose variation and makes good performance in testing data set.Besides, this paper try to explore an mapping between the frontal face image and non-frontal face image by using deep belief networks, by relaxing the limitation that the output are absolutely equal to the input, which just guarantees the output and the input are in equivalent in a high-level. Experimental results show the deep belief networks are capable of mapping the holistic information of frontal face and non-frontal face, but neglecting some details and diversity. In the paper, we employ deep belief networks by preserving the class neighborhood to classify the pose, which leads the images stand close to a few of those belong to the same pose. This approach performs pretty well in classifying the image by pose, but the feature doesn’t work in face recognition, since the difference between individual is losing in the learning procedure.This paper is a research on face recognition based on deep learning. Beside the approach mentioned above, we seek the potential applications based on deep belief networks.