Research On Small Sample Face Recognition Based On Convolution Neural Network

With the acceleration of information technology,people have higher security requirements for information.This also brings a great challenge to the research of face recognition.These challenges mainly come from facial gesture,facial expression,illumination and occlusion.The traditional method is difficult to achieve an ideal recognition rate.Recently,with the development of convolution neural networks.The convolutional neural network has strong robustness toface recognition.The drawback of it is that the network structure is huge and the training weight is large,so it needs massive training data to train,while due to the limit of storage space and sample time,there is often lack of traning samples,it will directly affect the accuracy of the recognition.In this paper,we propose a novel face recognition method based on convolutional neural network and traditional feature extraction.The main work is as follows:Firstly,in the preprocessing phase,this paper normalizes the face image to a uniform grayscale image,and then use the image rotation and translation to construct virtual samples.One image is transformed to produce 17 training images,which can solve the small sample problem to a certain extent.Secondly,in the feature extraction phase,three different methods,Discrete Wavelet Transform(DWT),Local Binary Pattern(LBP)and Sobel Operator,are used to extract the low-frequency Sub-map,LBP texture feature map,Sobel edge feature map separately.These three kinds of features are complementary to each other,which lays the foundation for the following feature fusion and classification.At last,to solve the problem of small sample face recognition,a single-level multi-scale convolutional neural network(smCNN)structure is designed in the feature fusion and classification stage.The network architecture consists of a convolution layer,two sampling layers,two fully connected layers and a Softmax layer.Convolution kernels of four scales are set in the convolution layer,and the three different feature maps extracted in the previous step are merged by multi-scale thought.The merged features are finally classified by the Softmax layer.Also,the network structure applies the ReLU,LRN,DROPOUT and other technologies to effectively prevent the over-fitting problem.To verify the effectiveness of this method,the AR,ORL and YaleB + Extended YaleB databases were selected in the experiment.The results show that the face recognition method proposed in this paper has strong robustness togesture,occlusion and expression,which is better than the current method with higher recognition rate.Although the recognition rate is slightly lower than that of LPA + ICI + mLBP method in the problem of illumination,but it also reaches a high level.In addition,since the number of samples used in this paper is small,it shows that this method has a significant effect on the small sample problem.