Research On Low Resolution Face Recognition

With the rapid development of computer vision,the biometrics technology field has made great progress.It is difficult to identify biological identities efficiently and accurately in recent years.Compared with traditional biometric identification such as fingerprint and iris,face recognition has the advantages of non-contact,easy acquisition,etc.It can be widely used in public security,finance,transportation,etc.Medical and other industries and fields.In order to more effectively apply face recognition technology in non-constrained environment such as video surveillance,we will study single sample and low-resolution face recognition in this paper.For low-resolution face images,which suffer from serious loss of detail information,deep learning based methods are used for feature extraction in this paper.The typical network in face recognition includes FaceNet,DeepID,VGG and so on.Because of the dismatch of features between low-resolution face image and high resolution face image,this paper adopts the re-trained model based on low-resolution face dataset after data cleaning,therefore,the extracted features are robust to facial pose and local occlusion,and which is discrimitive to different faces.Because a large number of samples are learned through the deep network,extracting the full connecting layer as a feature descriptor has more essential and robust image features,which can effectively improve the subsequent recognition effect of low-resolution images.Face recognition using a single sample per person in the gallery is really common in the surveillance system,we refer this as single sample problem,which means limited number of samples can be accessed,the classifier cannot be trained in a supervised learninsg manner.At the same time,each subject contains only one image.In this regard,we propose a low-resolution face recosgnition algorithm based on a sinsgle sample in this paper.Firstly,Cextracting the features of target images and identified images,then two methods are used to expand the registration sample respectively.Similarity based sample augment method is based on the principle that the high-resolution image features of the same target subject are larger than those of different target subjects,which acquires the threshold based on statistic of non-tarsget samples,then the identified samples are extended to the target subjects when the similarity degree falls within the confidence range.Image transformation based sample augment based method extend the targets subjects by the way of translation,slight rotation and local small random occlusion.On this basis,the machine learning method is used to construct the classifier and complete the single sample and low-resolution face recognition.