Research On Face Liveness Detection Algorithm Based On Multi-mode Fusion

Face liveness detection have become increasingly important with the popularization of face recognition,It aims to determine whether the face in front of the camera is a real face or a photo or other spoofing face.However,the diversity of face spoofing methods puts forward higher requirements on the face live detection system.Most of the existing face detection technologies use a single feature to judge,so their accuracy and robustness need to be improved.In addition,under the premise of ensuring the performance of the model,the computational efficiency of face liveness detection is also increasingly important with the expansion of face recognition applications.For the above problems,this paper proposes a multi-mode fusion face live detection technology and conducts an in-depth research.Based on the previous work in related fields,the main research contents and innovations of this paper are listed as follows:(1)In feature extraction,considering the limitations of the RGB color space in image analysis and the effectiveness of dynamic features in face liveness detection,feature extraction is performed from full face and local patch images on the HSV and YCb Cr color spaces and temporal images with dynamic information.In order to efficiently use the feature information of the three modalities,we propose to use the early fusion method to perform feature fusion,and design a multi-input fusion network.For the deep feature vectors obtained after fusion,this paper combines a residual structure to design a decision network with the combination of three different hierarchical features of high,middle and low to iteratively train it,and improves the original residual module in the decision network.By adding a 1 × 1convolution operation to enhance the expressive power of the network.Finally,the test was conducted on three public data sets.The test results showed that the proposed multi-mode fusion method is superior to the existing face liveness detection methods.(2)For improving the speed of model by optimizing the face live detection mode.In this paper,network slimming and model compression methods based on channel pruning are mainly used to optimize network model parameters.Network slimming is mainly achieved by simplifying the input and simplifying the structure.In terms of input simplification,we separated the images on the channels that can provide more discriminative features throughexperiments,and then used the image channel fusion to obtain the input images of the network.For the simplified structure,the network parameters are reduced by reasonably compressing the network structure and appropriately reducing the number of feature maps.Finally,it is proved by experiments that the model parameters are further compressed under this scheme,the model speed is greatly improved and the model still has a high accuracy.