Research On Face Anti-spoofing Models Based On Lightweight Network And Multi-source Feature Fusion

Face recognition has always been the first choice for personal identification,but most face recognition systems are vulnerable to forgery attacks by others,thus causing property loss,identity impersonation,and other risks to users.Therefore,how to efficiently identify spoofing faces is an urgent problem in this scenario.Researchers have mainly used deep learning-based multi-modal methods to achieve spoofing face detection,which is more efficient than single-modal methods.However,considering that in practical applications,not all occasions have the ability to collect images of various modalities.Moreover,deep learning-based methods usually have a large number of parameters and calculations,making model deployment challenging in scenarios with limited computing power.To address the above issues,based on designing a generalized lightweight feature extraction network and data enhancement strategy,this thesis carries out a series of endto-end face anti-spoofing methods around the modal data that can be obtained for different application scenarios,with specific work as follows.(1)Aiming at the problems of large parameters and high calculation cost of deep learning method,a face anti-spoofing method based on lightweight feature extraction network is proposed.In terms of the model,a lightweight feature extraction network(LFENet)with strong versatility is designed,which alleviates the problem of feature map redundancy through a depthwise center difference convolution-based ghost module(DWCDCGM).In the test phase,a data augmentation strategy based on random regional blocks is proposed.The experimental results show that the method achieves 0.5155%Average Classification Error Rate(ACER)on the CASIA-SURF(Depth)dataset,and model parameters are only 0.3445 M.(2)For the application scenarios where only single-modal data can be obtained,a face anti-spoofing method based on wavelet transform and two-stream convolutional network is proposed.The method transforms the form of the image by wavelet transform to enhance the detailed information of the original image,and then constructs a twostream convolutional network for multi-source feature extraction and fusion with LFENet as the backbone network.Finally,the pruning method was used to remove the redundant weights from the model and accelerate the inference time.Also experimented on CASIASURF(Depth)dataset,the ACER is reduced by 0.1353% with the same number of parameters.In addition,the above two methods also achieved good experimental results on four datasets,including SIW and OULU-NPU.(3)For the application scenarios where multi-modal data can be obtained,a face anti-spoofing method that cross-fuses multi-modal features is proposed.The method also uses LFENet as the backbone network,but constructs three branches to extract features from different modalities.Then cross-attention is designed to let the modality that works best when trained in a single branch to assist in generating attention maps for other modalities and is used for multi-source feature fusion.Experimental results show that the method achieves an ACER of 0.0578% on the CASIA-SURF dataset,which is 0.3224%lower than the two-stream network-based method.In addition,the method also achieves good experimental results on seven protocols of the WMCA dataset.To sum up,the three methods designed in this thesis can identify face spoofing attacks with high precision,among which the multi-modal method works the best,and the number of parameters is only 1.0607 M,which is much lower than the existing multimodal methods.In practical applications,different models can be selected for face spoofing detection according to the ability to obtain data in different application scenarios.