Research On Face Anti-spoofing And Recognition Algorithm

Face recognition is a very important field in computer vision.At the same time,it is widely used in the field of social life,but there are security risks caused by spoofing faces.Face anti-spoofing is proposed in recent years to cope with the problem of spoofing faces,and it is important for improving the security of face recognition systems.From the traditional machine learning to the convolutional neural network,the face recognition algorithm and the face anti-spoofing algorithm gradually improve the performance in the development.Currently,the convolutional neural networks based are the mainstream in the development of the two algorithms.However,the two algorithms are usually considered as separate tasks,ignoring the correlation between the two tasks.At the same time,the face biometric detection algorithm generally suffers from insufficient generalization.By analyzing the correlation between face detection and face recognition tasks and the generalization of face detection algorithm,a face detection and recognition algorithm based on multi-task learning and metric learning is proposed.In this algorithm,a multi-task convolutional neural network for face detection and face recognition is designed.At the same time,in order to further improve the generalization of face anti-spoofing,the idea of metric learning is used.A Stochastic Pairwise Confusion Loss(SPC-Loss)is proposed for training the branch of the face anti-spoofing network.The final algorithm model is obtained by training the multi-task convolutional neural network using the SPC-loss and classification loss functions.Experiments show that the algorithm has excellent performance in face anti-spoofing and face recognition,and it is more generalized than other algorithms in inter-test for face anti-spoofing.In order to narrow the difference in the distribution of the face anti-spoofing data and improve the generalization of the face anti-spoofing algorithm,a Domain Adaptation algorithm is proposed,trying to transfer face anti-spoofing data into one domain.The algorithm uses a pre-trained convolutional neural network to extract features,and designs a content loss function and a domain loss function by measuring two image features.A feedforward neural network is trained as an image conversion network to achieve the purpose of transfer samples to one domain.Experiments show that the algorithm can reduce the feature differences between the two domain of data and can be used to improve the generalization of the face anti-spoofing algorithm.