| In terms of identity authentication technology based on biological feathers,facial recognition enjoys wide attention and is regarded as a highly challenging area.Since many studies in related subjects such as machine vision,image processing,stereopisis,machine learning deepen continuously,facial recognition is widely-used in people's daily life,such as access control,unmanned supermarkets,face payment and airport security checkpoints etc.Although traditional face recognition system can identify different faces,it is difficult to determine whether the face is a living body,a photograph,or a model.Therefore,the appearance of human face detection technology is precisely to eliminate the hidden dangers in the above-mentioned face recognition technology,so as to make sure the stable operation of face recognition system.The thesis summarizes previous research results of human face detection both at home and abroad.It focuses on photo-spoofing and face model attack problems.It studies the optical flow,fusion depth and texture disturbance information,and combines Support Vector Machine(SVM),designing two types of methods to attack faux-face.The main work of the paper is as follows:1.Improved human face detection based on optical flow method.By using a 940nm narrow-band infrared camera with 940nm infrared light to capture face images,it analyzes the different characteristics of face photos and real faces under light flow,then classifies them by SVM.In this way,the detection accuracy under a single optical flow chart is improved to 83.3%.Positive sample detection rate was 82.8%,while negative sample detection rate was 83.9%.Both are more accurate than a color camera under the same environment.2.The live detection method by combining depth information and texture disturbance information is proposed for the first time.The depth information of human face is acquired by using the Intel Real-sense D435 camera,then the differences between three-dimensional features of the true and false human faces are analyzed.In order to prevent face models with depth information attacks,this article simultaneously obtains face color image information,the texture disturbance features from non-rigid changes of human faces,the method can achieve live body discrimination in the absence of user cooperation,with an accuracy rate of 98.5%.positive sample detection rate of 98.9%and negative sample detection rate of 98.0%respectively.3.Two face databases are constructed.One is a 940nm narrow-band infrared camera video database,including 30 living human faces and 60 human faces photos.The optical flow live detection method proposed in this paper was verified experimentally on the data set.The results show that this technique can effectively distinguish between true and false faces.Another face database with depth information recorded by Intel RealSense D435,including 30 live faces and 60 photo faces,approves that the method of integrating depth information and texture disturbance information can effectively solve the problem of living body detection without the cooperation of paticipants.4.Using the Up Board and D435 depth camera to establish a set of human face recognition system based on live detection. |
