Liveness Detection Of Occluded Face Based On Dual-Modality Convolutional Neural Network

Face liveness detection is an important part of the face recognition.The existing studies do not consider the problem of the scene with occluded face in liveness detection.In this thesis,to solve the problem that the partial area of the face can be detected decreases and the detection accuracy is affected when the face is occluded.We proposed a detection scheme based on dual-modality feature fusion.The main work of this thesis is as follows:We built a dual-modality model for occluded face in face liveness detection.The residual network is used to build the dual-modality convolutional neural network.To make the face liveness detection under the occlusion can get more image information,we use natural light face gray image and near infrared face gray image as a dual mode to detect the human face living.as testing data.We designed a single,dual-modality face liveness detection experiment for comparison.The experimental results show that under the occlusion degree of 10%-20%,the accuracy of the dual-modality detection scheme has a lot of improvement,which can reach 96.0%-97.4%and 97.0%-97.93%respectively.The advantages of the dual-modality face detection model in face occlusion are verified.2.Improving the loss function.To fuse the model network structure and image description algorithm,the model loss function was improved based on the dual-mode residual network In view of the characteristics of good attack image reshooting technology and high similarity between living and attack images,structural similarity algorithm is used to improve the loss function,so that the model training can reduce the error between the predicted value and the label value,and put the distance apart between the living and attack face image features.Subsequently,To improve the feature extraction quality of the single-modality face liveness detection model,the single modality feature extraction and fusion method were improved.The comparison experiments of single-modality face liveness detection before and after the improvement of loss function,and the experiments of dual-modality face liveness detection before and after the improvement of feature extraction and fusion scheme are designed respectively.The experiments verify that the improved loss function and feature extraction method can improve the accuracy of model detection to some extent.Finally,the liveness detection accuracy reached 97.6%with 40%occlusion degree.3.We built a face liveness detection platform.The platform embedded the face liveness detection model designed in this paper into the system.The platform embedded the face liveness detection model built in this thesis into the system,providing users with a graphical interface for face liveness detection.The platform is a Web system that can be accessed through a browser.The platform has three modules,which are the user,the administrator and the face liveness detection part.The user uploads the image according to the prompts and carries out the face liveness detection.Administrators can view and query background detection records.The visual face in liveness detection platform improves user experience and facilitates data management by administrators.