Face Verification Based On Deep Learning

Face verification is to judge whether two face images are the same person.Compared with other identification methods, such as ID card, password, fingerprint recognition and iris recognition. Face verification has the advantages of not easy to be stolen, strong stability,hidden operation and non-contact acquisition etc. So face verification can be widely used in security monitoring, public security, e-comerce and other fields. However, in practical applications, the face images collected by the camera often contain illumination, facial expression and pose interference, which can severely reduce the accuracy of face verification.Therefore, based on the excellent features extraction ability of deep learning theory, in this paper we study to extract the feature of a facial image with illumination, facial expression and pose interference, to make computers can be accurately identified. The main work is as follows:(1)Present a deep reconstruction networkDue to the interference of illumination, expression and pose in face images, the accuracy of face verification is reduced, while the traditional image feature extraction methods only robust with one or both of them. Our algorithm has two advantages. First, the traditional deep learning algorithms such as denoising autoencoder learning to get the ability to reduce Gauss noise or other image noise. In this paper, the algorithm is designed to remove the interference of the face image, such as illumination, expression and pose. Therefore, the algorithm in this paper is more targeted to the illumination, facial expression and pose interference, and can effectively remove these three kinds of interference. Second, the algorithm can extract the image feature after removing the interference. In this way, we can avoid the information loss caused by the feature extraction of the reconstructed face. The experimental results on CMU-PIE database show that the our algorithm can effectively remove the illumination, facial expression and pose interference. And the extracted robust face image features.(2)Face verification using deep neural networkIt is a similarity measure to judge whether the two facial features belong to the same person. However the face features are complex and the vector dimension is high. So,the traditional methods such as support vector machine (SVM) and K mean clustering can not be used to measure the features of the human face. Because the deep neural network has excellent nonlinear mapping ability and model expression ability, so this paper use deep neural network to learn a effectively reflect implicit data similarity measure function from thirty thousand groups of training samples. The experimental results on CMU-PIE database show that this neural network can achieve high accuracy by using the proposed method.