Research On Face Recognition Based On Gabor Convolutional Neural Network

Face recognition,which stands out from various kinds of biometric identification technology due to its friendliness,directness and high acceptability,is used more and more widely in practical application.With years' development,face recognition has made tremendous progress,but its accuracy can be easily affected by many objective factors,such as varying illumination,pose,expression and partial occlusion.So how to further improve the recognition rate under natural scene is still a challenging problem to be solved.At present,most conventional face recognition algorithms are feature-based methods,in which prior knowledge is needed to manually extract the shallow facial features and these features are then fed into a suitable classifier.Deep learning,which has a hierarchical structure inspired by human brain's multiple stages of information processing,and can perform automatic high-level feature extraction and classification in a module even under complex conditions,has gradually been a research hotspot in the field of pattern recognition.Among these deep learning models,Convolutional Neural Network(CNN),as a multilayer perceptron,adopts the idea of local receptive fields to ensure some degree of invariance of shift,scale and distortion of input 2-D picture's,therefore it is widely used in face recognition.But the number of training samples is likely to be far less than the number of parameters in CNN,which will result in a serious over-fitting.Besides,the training time will get longer with the increase of the depth and scale.In order to solve the above problems,while keep the depth of network,a Gabor Convolutional Neural Network(GCNN)is proposed for face recognition.Compared with the classical CNN,GCNN improves the network structure and learning algorithm as follows:(1)The convolution kernels in C1 convolution layer are replaced by a series of Gabor filters,which simulate simple cells in primary visual cortex V1,to extract elementary visual features;(2)Maximum operation is applied to two feature maps obtained using filters with neighboring size and same orientation from G1 layer,in S2 pooling layer;(3)The connecting relationship between S2 pooling layer and C3 layer is reconstructed according to the prior knowledge;(4)Dropout technology is introduced into the Mini-Batch Stochastic Gradient Descent algorithm to improve the generalization ability.To demonstrate the feasibility of GCNN on the small sample face recognition tasks,we apply it to ORL and AR face database.The first three experiments focus on the influence of the weight initialization,the connecting relationship between S2 pooling layer and C3 layer,and the number of neurons in fully connected layer on the recognition rate of GCNN.The middle two experiments compare the recognition rate of GCNN under certain parameters with that of other general face recognition algorithms.Experimental results show that the improved GCNN exhibits better generalization ability and saves more training time than the classical CNN;On the other hand,the proposed GCNN outperforms existing works in achieving higher recognition rate on both ORL and AR face database.The last experiment validates the applicability of GCNN in face recognition under natural scene.This method of face recognition has a certain theoretical significance and practical value.