Facial Feature Extraction And Matching Based On Deep Learning

With the rapid development of computer vision and pattern recognition and the increasingly demand of human feature recognition in society, voiceprint recognition,fingerprint recognition and constrained face recognition can no longer meet people requirements. With the increasing requirements for unconstrained face recognition,some world’s largest internet companies have already started research work in this area.And extracting robust features from face images and having a good classification algorithm to separate them both play vital roles in unconstrained face recognition.On the basis of the theory of deep learning, this thesis do research on feature extraction and feature matching steps of face recognition and implement the algothrims.The main work of this thesis include the following:1. Summary the conventional methods of face feature extraction and matching.Introduce the concept and the development of Deep Learning by studying the literature.Deep Learning is a machine learning method and it’s mainly used for research of large data set and can learn the deep feature representation of the data.2. I study on the deep convolutional neural network in this thesis, then improve its network structure loss function and design the feature extraction network based on the dataset gathered in this thesis, so that it can extract deep facial features from face images and prevent the occurrence of over-fitting. Such features of the human faces have a better representation than artificial design features and features extracted by ordinary convolutional neural network so that using such features can achieve better results in the feature similarity matching step.3. In the facial feature matching step, I propose a step-matching algorithm which is divided into two steps:fast matching and fine matching and the purpose of the algorithm is to shorten the time cost by feature matching. I also choose different learning algorithms as the algorithm of fine matching step in this thesis to compare their effects on the recognition rate of face recognition. Eventually Joint Bayesian is chosen as the fine matching algorithm in this thesis.4. In this thesis, I implement the improved deep convolutional neural network and the step-matching algorithm.To solve the problem of taking too much time in training step occurred in the experiment, I reimplement the network, using GPU training methodinstead of CPU training method in this thesis to train the network to achieve the purpose of speed up training.