A Study Of Face Recognition Based On Convolutional Neural Network

With the vigorous development of information technology,great changes take place in both people’s learning and life style.At the same time,public security has increasingly become the focus of attention.Biometric identification technology emerges under such circumstances,especially face recognition.Compared with other biometric technologies like fingerprint recognition and iris recognition,face recognition takes the advantages of its friendly acquisition mode,less needing or no needing of cooperation,simple operation and good imperceptibility.Face recognition has wide application prospect in the field of public safety,financial security,etc..It has been a hot research topic in the security technology field.The rapid development of computer technology makes the study of face recognition increasingly intensified.As well known,the factors such as occlusion,illumination,and expression are longterm challenges that face recognition has been facing.This urgent needs to build a description model that is insensitive to these factors.The appearance of deep convolutional neural network has greatly improved the status of face recognition.Compared with traditional recognition approaches,convolutional neural network(CNN)needs no complex and timeconsuming feature extraction by manual,but can perform automatic feature learning with a powerful network framework.At present,it is one of the most popular method for face recognition,and can cope well with the rapid accumulation of the vast amounts of face data.Its successful applications have been achieved in the real scene.This thesis aims at improving generalization performance for module and extracting more effective features,with no need of increasing data.This dissertation discusses the issues based on CNN as follows:(1)Propose a patch strategy for deep face recognition.The critical points of the method are to embed the patch strategy into CNN for complementary and efficient face representation.Specifically,a new network layer is created to implement operation of cropping patches in CNN models,and a multi-branch CNN is also constructed to learn and fuse patch features simultaneously.The multi-branch network structure realizes online patching,multi-feature extraction and feature fusion in an end-to-end fashion.The features extracted by CNN are global when taking a whole face image as input.This kind of global features is discriminative but not effective enough due to its sensitivity to local variations.A simple and direct access to more efficient features is to introduce local features,which pays more attention to encode the detailed attributes within the specific area.The introduction of local features can make up for shortcomings of global features.Since the patch strategy is an available way to obtain local information,the multi-branch network based on the patch strategy is proposed.The proposed approach uniformly samples six image patches of the same size,according to five sparse facial key points(two eye centers,the nose tip and two mouth corners).For each patch,it is sent to a network branch to learn and extract its feature.Further,all features are normalized and fused to form the final face representation.Compared with existing multi-patch based CNNs,no extra space is needed to store the local patches since the method takes an entire face image as input and the images are cropped online.More importantly,the method can promote the interaction between local information and global information and that among local information,so that the features can be weakened or strengthened adaptively.This is because the parameters of each patch are optimized by the end-to-end training,and thus the face representations are intensified.Experimental results on the LFW and YTF datasets show the proposed system achieves comparable performance with other state-ofthe-art methods with less training data.Especially,it is quite suitable for the problem of occlusions,poses,expressions and illuminations.(2)Propose a regularization technique to optimize CNN by kernel-based decorrelation.The problem of overfitting often brings bad performance of CNN,which can be alleviated by a valid regularization.For the model presenting overfitting,there is a lot of redundancy in parameters.The redundancy parameters usually capture similar patterns and have strong correlations.The proposed method attempts to make use of kernel functions to regularize CNNs by reducing the correlation of weights,thereby ensuring maximization of compiled information.In contrast to decorrelation on features,which is usually in a high dimensional feature space,the proposed method directly performs this operation on weights.Thus,the computational effort can be greatly reduced.As well known,kernel function is defined as the inner product in the feature space,so it can be used to measure the similarity between features.By exploiting Gaussian kernel function to measure the correlation of weights,this algorithm not only takes advantage of the angle of weight vectors,but also considers the effect of the distance between them.The larger the value of Gaussian kernel,the greater the correlation between two weight vectors.Especially,an adaptive estimation of kernel size is used to update the width of kernel in each iteration when training CNNs,avoiding the disadvantage of the sensitiveness to it.The method exhibits good compatibility with other regularization such as Dropout and BatchNorm,and more improvements are achieved when they are performed together.Additionally,it is widely available for various networks of different capabilities.Extensive experimental show that the method performs better on several datasets for both object recognition and face recognition.Compared with other regularization based on decorrelation,the proposed approach is superior,especially to the large scale face recognition.