Research Of Face Liveness Detection For Mobile Terminals

Face recognition is gradually applied in areas such as security,financial and personnel management.The liveness detection,which determines whether the recognized face is a true face from a live human or not,is a security guard of face recognition system in these important applications.Hence it has attracted more and more attention in recent years.With the development of mobile Internet,the main application field of face recognition and liveness detection also gradually transferred to the mobile terminal whose main representative is the smart phone.Because of the complexity of liveness detection problem itself,mobile terminal hardware limitations and complex mobile application environment,to accomplish precise and reliable liveness detection in mobile terminals faces many difficulties.In this paper,according to the characteristics of application background of liveness detection of the mobile terminal,liveness detection database is established.At the same time,image preprocessing,liveness feature extraction and classification,and liveness detection algorithm are researched.Some explorations are done for convenient,rapid and accurate face liveness detection in the mobile terminal.In image preprocessing,the judgement of light intensity reduces mobile lighting unstable factors;Distance judgment from the camera is to obtain ideal shooting range;Face alignment can reduce facial gesture factors;Image equalization can abate lighting effects.According to the characteristics of the mobile terminal application,simple LBP,HOG and Haar liveness features are selected and SVM and Adaboost classifiers are selected.Three different schemes for mobile terminal are studied which are traditional comprehensive interactive liveness detection,secondary imaging liveness detection and multitask deep learning liveness detection and their characteristics and test results are compared and analyzed.Three schemes have ordinary single camera in common conforms to the mobile terminal characteristics.It does not need to neither specific camera nor multiple cameras.The traditional interactive liveness detection is based on randomly combine many liveness properties which contains the blink,open mouth detection,detecting shook his head,micro mouth detection and covered eyes detection.Random combination of multi liveness property can effectively prevent cheating video ready and enhance the capacity of anti attack.At the same time,the scheme is high feasibility,a small amount of calculation and strong adaptability and accuracy rate is 98% on testing set.Secondary imaging liveness detection is mainly based on the principle of micro texture image lost when the secondary imaging.The biggest advantage is convenient to use and users don't need to cooperate.To reduce the burden of mobile computing decreases feature dimensions by Adaboost algorithm.The accuracy rate is 92% on the testing set.Multitask deep learning liveness detection is the improvements on traditional comprehensive interactive liveness detection.To solve the problem of multiple independent model and many times of feature extraction in the tradition model,multitask deep learning network model learns the joint characteristics of many liveness properties and guarantee liveness detection accuracy.