Face Anti-spoofing By Using Feature Fusion

Face anti-spoofing is an important derivative subject of face recognition research.Due to the wide application of face recognition in daily life,the research on face anti-spoofing has attracted the attention of many researchers.To effectively distinguish between real faces and fake faces and improve face recognition accuracy,researchers have explored this problem from different angles.In the early days,limited by the lack of data sets,the research could only be carried out from the aspects of texture and image quality;with the continuous improvement of related data sets,it is now possible to research other perspectives such as depth and illumination.However,the continuous development of face anti-spoofing technology has also led to the increasing number of fake types,making how to obtain the real and fake two more representative face features have become one of the current research focuses;How to obtain more representative facial features of real and forged is one of the current research focuses.This article starts from the difference between a real face and a fake face,Propose a face anti-spoofing structure,the main research contents of this article are as follows:1.To effectively distinguish between real faces and fake faces,this research combines deep learning features and ICA features in the feature extraction stage;deep features use ResNet combined with SE-Net network to obtain fused features,which are effectively enhanced the Correlation of each feature channel,the features extracted by the independent component analysis ICA method acquire features that contain rich high-order information,then the two parts of features are spliced and merged,SE-Net network is used again for feature fusion;finally,the obtained features are classified into real and fake.Perform experimental verification on the data set.The results show that a better classification accuracy rate has been achieved through experimental comparison with other related excellent algorithms.2.To improve the generalization ability of the method,the regularization method is added to improve the balance ability of the network and the loss function is improved.It is compared with the traditional regularization method.Firstly,the regularization method is applied to the SE-ResNet network,and experiments are carried out.Several groups of parameters are selected for experiments,and the parameters with the best effect are adopted.Secondly,the improved network with the best performance combined with the ICA method is compared with many other excellent algorithms.The comprehensive evaluation shows that the introduction of this regularization method makes the feature extraction of the backbone network more balanced and more capable of characterization.