Research On Face Liveness Detection Based On Deep Learnin

With social development and technological progress,the face recognition technology has stepped into the human life and has a profound impact on their life style.Though the face recognition system brings lots of convenience,great security risks arise in its application,for example,photos displayed by mobile phones or tablets would make the face recognition system mistakenly match the forged face with the existing face identity information in the database and give permission,and therefore low security,whose reason is because of no attention that system pays to check the face collected for recognition living or fake.Face anti-spoofing detection can distinguish the image is from a living body or not.The fake attack can be blocked to avoid wrong judgment,which therefore improve the system security.To deal with the aforementioned problems,this thesis studies the face anti-spoofing detection methods as follows.The comparison between the fake and live face images show the gradient information plays an important role in their feature extraction,since their edge details are visibly different.To make fully use of these features,the directional differential convolution(DDC)is proposed through detailed analysis on various convolution operators.Its combination with traditional convolution can be balanced by parameter θ to enhance convolution.To verify its effectiveness,the CDCN and CDCN++ networks are introduced for testing via CASIAMFSD,replay-attack,OULU-NPU and MSU-MFSD datasets,respectively.Experiments show that the combined convolution could strength their two advantages in most cases,which improve the feature extraction ability of convolutional neural network and the robustness of face anti-spoofing detection.Transformer in Transformer(TNT)is an image multi-classification model proposed by Huawei Noah Lab,which divides the image evenly and then divides the obtained blocks evenly again.Based on self-attention mechanism,it can encode and decode the position and semantics of sub-block,which could extract the salient features and the position information of each image to fuse into coding sequence.Its fully extraction of typical features from image is greatly relevant to the face anti-spoofing detection.Given its successful application in image processing,TNT model is firstly introduced for the face anti-spoofing detection in this thesis.By adjusting the number of image segmentation and network outputs,the model accuracy is superior in most subtypes of CASIA-MFSD,Replay-Attack and MSU-MFSD datasets.In OULU-NPU dataset,its accuracy is no less than that of the combined convolution in convolutional neural network,and leads in several sub-protocols.Experiments show that TNT model could distinguish subtle differences among images and has a good feature extraction ability,which provides a new idea for the face anti-spoofing detection in future.