Research And Implementation Of Face Anti-spoofing Based On Multi-color Space

Face anti-spoofing(FAS)can solve the security problems caused by forged face attacks,and has been widely used in identity authentication systems.Face anti-spoofing based on single frame image can not use dynamic information and multimodal information,and it is difficult to resist the influence of illumination,scene and other environmental factors,resulting in poor generalization ability of the algorithm.By analyzing the difference between real faces and forgery attacks in multi-color space,this paper designs a network model structure of adaptive color space extraction and multi-dimensional color space feature fusion,realizes the information expansion and multi-dimensional fusion of the original image data at the color space level,and improves the generalization and accuracy of the algorithm.Aiming at the problem of poor generalization ability of face anti-spoofing algorithm based on single frame image in complex environment,a face living detection algorithm based on adaptive color space extraction is designed in this paper.By analyzing the transformation formulas between different color spaces,the algorithm maps the transformation relationship into the convolution neural network,and makes the convolution neural network adaptively extract the color space data? In order to maximize the use of the complementary information of multi-color space under different environmental conditions and improve the generalization of the algorithm,the attention model is introduced to allocate the weight of color channel features and image space features,and the depth features with color and texture discrimination are obtained by weighted fusion.Experiments show that the ACER of the proposed algorithm is reduced by 2.8 % compared with the benchmark model.Aiming at the problem of low accuracy caused by the lack of feature fusion across color space dimensions in existing neural networks,a multi-dimensional feature fusion face antispoofing algorithm is designed in this paper.The algorithm designs a multi-dimensional convolution module based on grouping convolution and combined convolution.Under the condition of constant parameters,the original tree neural network is expanded into a mesh neural network,which increases the ability of network information interaction and solves the problem of lack of cross-channel information fusion in the multi stream CNN network structure.Experiments show that the average classification error rate of the proposed algorithm is lower than that of other classical combined convolutional networks,and the HTER is lower than that of other face living detection algorithms using multi-color space.Based on the proposed face detection algorithm,this paper designs a face anti-spoofing system based on RGB camera.The system has the functions of face recognition,face early warning and face anti-spoofing.It can run stably and efficiently in complex environment and meet the relevant hardware and scene constraints.At present,it has been actually deployed in relevant projects.