Face Anti-spoofing Detection Based On Convolutional Neural Network

As humanity enters the digital age,more and more scenes in daily life require people to authenticate their identities.And face facial features due to its advantages of non-contact and stability in many identity authentication way occupies a very important position,in the face recognition and validation of human face tracking,monitoring,facial expression recognition,facial attribute recognition,facial illumination adjustment and deformation,image video retrieval,and other fields has been widely used.Today,it’s easy to fake someone’s face every day through video or even surveillance footage.Once the face recognition system is deceived by forged face,it will bring very bad influence to individuals and society.Therefore,face anti-spoofing detection has become an important research direction in the field of computer vision,attracting many scholars worldwide.At the same time,the emergence of deep convolutional networks provides a new idea and method for face spoofing detection.Face anti-spoofing detection using convolutional neural network is the main content of this topic,the specific work is as follows:1.Literature survey found that the current face anti-spoofing field is faced with complex attack means,changeable scene lighting,less data and other problems.Existing can achieve high precision high generalization ability of network mostly large and heavy network is adopted,mainly through the number of deepen the depth and the characteristics of the network to obtain the network model of performance improvement,although this can achieve high precision,but very tall to the requirement of hardware to calculate force,cause the slow training speed,high difficulty.Moreover,the most important application of face anti-spoofing detection is to be deployed in the camera terminal as a front module of face recognition,and the computing power of the terminal equipment is very limited,so the heavy network will not be the future development trend of this field.How to improve accuracy with less hardware computing power has become a pain point in this field.2.In view of the problem that there are few light and small network solutions and low accuracy in face anti-spoofing field,Mobile Net V2,a lightweight network,is selected as the backbone network of the model in this topic.Mobile Net V2 uses bottleneck convolution block and reverse residual structure to optimize the computation and spatial complexity of the network.It has certain advantages in many lightweight neural networks and provides a good foundation for the research of the subject.In order to improve the original network,the original convolutional layer of Mobile Net V2 is replaced by central differential convolutional layer,and Mobile＿CDCNet A and Mobile＿CDCNet B are proposed.From the perspective of adding spatial location features,the bottleneck convolution block is improved by using coordinate attention mechanism,and Mobile＿CANet is proposed.The three models improved by 2.5%,4.7% and 4.2% on average compared with the original network on CASIA-FASD,OULU-NPU,Replay-Attack and MSU-MFSD data sets.3.From the perspective of multi-scale feature fusion,the central difference convolution and coordinate attention mechanism are simultaneously integrated into the original Mobile Net V2 network,and Mobile CDCNet is proposed.The accuracy of Mobile CDCNet in four data sets reaches 95.9%,93.4%,99.9% and 98.8% respectively,which is the highest among the four models.In addition,it also has good performance in cross-dataset experiments.Compared with the existing methods,this model shows lower complexity and higher recognition rate,which helps to solve the pain point problem in this field.