Face Recognition Based On Deep Learning

Face recognition technology has a lot of advantages including collecting data easily,high speed and precision of validation.Therefore it has received widespread attention from more and more people and been used in scientific research,finance,security,Internet and so on.With the development of deep learning,especially the convolutional neural network,face recognition technology has also got rapid development.Now face recognition technology has been gradually out of the lab and came into our lives.Thus it can be seen that the deep learning is very important to face recognition.It is studied that face recognition method based on the deep learning in this thesis.There are a number of improvements and innovations base on the existing algorithms are put forward in the paper and get three kinds of face recognition algorithm base on deep learning finally.At present the research results have obtained are as follows:Firstly,it is studied that how to improve the robustness of light of convolutional neural network.The local binary pattern(LBP)feature in traditional face recognition method is not sensitive to illumination changes,so this thesis proposes a face recognition algorithm base on LBP feature combined with convolutional neural network.The predominant feature of the algorithm is that changes the typical format of input data to convolutional neural network.Most often,the input data of convolutional neural network is original natural image or gray image,however,it is different in the algorithm.The first way is converting the original gray image to LBP feature map,and then fuse the gray image with LBP feature map as the new input data of convolutional neural network.Fusion of data contain more abundant information than the original data,so the algorithm indeed has a lot of improve robustness to illumination change after experimental verification.Secondly,the methods of deep learning model fusion are studied in this thesis.The core idea of boosting algorithm from traditional machine learning is that the multiple weak classifier will be combined into a strong classifier according to certain strategy.Influenced by the boosting algorithm,a new method,face recognition algorithm using the way deep learning model fusion base on Maxout,is put forward.The basic idea of this algorithm is to make the two convolutional neural network models that are poor performance combine a new network in the form of Maxout in the selected layer to make combined network hasthe advantages of the two simple network at the same time and reduce the shortcoming of the two models,so as to improve the performance of the whole network.The experimental results show that the fusion model is better than two simple network performance.Thirdly,there are some work has been done to study and improve the residual network.The residual network is composed of original residual units that use an important technology called identity mapping,but the identity mapping may not be the best choice for shallow convolutional neural network.So,a weighted residual unit automatically is offered and a hybrid residual network combined the automatically weighted residual units with original residual units is put forward in this thesis.Finally,a best way of mixing is selected through a variety of analysis.The experimental results show that the hybrid residual network is better than the network using only the original residual units or automatically weighted residual units.