Research On Face Recognition With Single Training Sample

Recently,computer vision is widely used in security and protection system.As face is convenient to be selected,face recognition technology has been one of the most popular technology in computer vision.But,in some special applications,eg: entry and exit administration,crime tracing,video surveillance etc.,there is always only one training sample,which promotes the research of single training sample face recognition.However,the existing papers mostly use traditional methods,which makes recognition rates very low.Deep convolutional neural network has a deep architecture,it can learn the high feature of image.Compared with traditional methods,it can extract deeper features.In the paper,it presents an expanding sample method for face recognition with single training sample;to validate the similarity between expanding sample and original samples,a method is proposed;to improve the quality of expanding samples,an expanding sample model is selected;to increase the rates of face recognition with single training sample,a method that merging traditional method into deep learning is put forward.The main content include:(1)In traditional method,to overcome the shortage of lacking training sample in single training sample face recognition,a novel expanding sample method is presented.According to the fact that different intra-class variation can be used in different individuals,it generates an intra-class variation set with the help of an extra data-set;single training sample and the intra-class variant set are used to expand sample.(2)Deep learning is used to do single training sample face recognition.Firstly,it uses transfer learning to bring in a deep convolutional neural network model which is well-trained and can express face image very well;Then,these expanding samples are used to fine-tune the model;Finally,the fine-tuned model is used to implement experiment.(3)To validate the similarity between expanding sample and original sample,a method is proposed.Firstly,the distance of corresponding pixel between expanding sample and original sample is measured;Secondly,the intra-class variant distance of same person same variant is counted,and its means is used as intra-class variant threshold;Then,the similarity of image is measured according to the threshold of intra-class variant;Finally,the wholesimilar images is counted,and the similarity between expanding sample and original sample is calculated.Expanding sample that with different similarity is used to implement experiment,the optical expanding samples are selected according to experimental result,at the same time,the intra-class variant set which is used to generate the optical expanding samples is selected as the optical expanding sample model which can be directly used to expand face sample combined with the expanding sample method.(4)A method for single training sample face recognition is proposed,which combined traditional method and deep learning.The method has the common advantage of traditional method and deep learning method.On one hand,it produces many training samples for deep learning method using traditional method which is convenient to be used;On the other hand,it can express and describe face image very well using deep learning method.Firstly,it brings a well-trained deep convolutional neural network which can express face image very well;Secondly,the expanding sample model is used to expand sample combining single sample and expanding sample method;Then,these expanding samples are used to fine-tune the network;Finally,the fine-tuned network is used to implement experiment.