Research On Face Liveness Detection Algorithm Based On Convolutional Neural Networks

In recent years,face recognition technology has become more and more popular,playing an increasingly important role in national security and social management.But the faces are exposed to the outside world,it is easy to be obtained by others.Especially some people with ulterior motives will attack the face recognition system with photos and the like.In order to improve the security of the face recognition,scholars have proposed the face liveness detection technology.Face liveness detection,which is to determine whether the face is a real face or a fake face,can be used to prevent the face recognition system from being invaded by illegal users,and to avoid unnecessary losses caused by deception by the attacker.It is an important security technology.Current mainstream face liveness detection algorithms have achieved good performance in multiple public databases by using deep learning technology,but they still have some problems.Because every database is different with each other in lighting conditions,background,etc.,the current methods have poor generalization performance in cross-database tests.In response to this problem,this paper studies the generalization performance in the crossdatabase experiments of face liveness detection based on convolutional neural networks,which has theoretical significance and practical value.The specific research work is as follows:1.A convolutional neural network for extracting color channel difference image features is proposed for face liveness detection.Firstly,by constructing color channel difference images to reduce the impact of non-spoofing information and highlight the spoof noise,the spoof noise information can better reflect the difference between real samples and attack samples.At the same time,an attention network is designed.By introducing the attention mechanism to optimize the network's learning of classification features,it ensures the model's ability to extract face fraud information in attack samples and improves generalization performance.Finally,the half total error rates of the cross-database testing experiments between the three public databases are between 34.5% and 36.2%,and the algorithm can improve the generalization ability stably and effectively.2.A dual-stream convolutional neural network for extracting the fusion features of RGB images and Gamma transform images is proposed for face liveness detection.RGB images contain rich detailed texture information,and Gamma-transformed images are more robust to lighting changes.In order to combine the advantages of the two images,a lightweight dualstream convolutional neural network for face liveness detection is proposed.The branch network extracts the features of the RGB image and the Gamma image respectively,and then obtain more robust classification information through feature fusion.The half total error rates of the cross-database testing experiments between three public databases are between 24.9% and 37.1%,indicating that the algorithm has good generalization performance,and the algorithm has fewer parameters of the model.