Research On 3D Face Reconstruction And Recognition Based On Static Image

Face recognition(FR)is a technology that uses face as biometric feature for identity recognition.Compared with fingerprint,iris and other widely used biometric technology,face recognition is more natural and has a broad application prospect.Although two-dimensional face recognition has made remarkable progress in recent years,its accuracy is very sensitive to posture,illumination,expression and occlusion.3D face recognition can make full use of reliable facial geometry information to overcome the above limitations.The 3D face detail geometry information determines the 3D face recognition accuracy.Obtaining the 3D face detail geometry shape is helpful to improve the 3D face recognition accuracy.Meanwhile,the 3D face recognition accuracy also needs to be improved under the condition of large posture and occlusion.Due to the large amount of 3D face data,the recognition speed is slow,under the premise of meeting the recognition accuracy,the current problem is how to use lightweight network to achieve real-time 3D face recognition.In order to solve the problem of 3D face detail reconstruction and 3D face real-time recognition based on lightweight network,this paper conducted in-depth research on 3D face reconstruction and 3D face recognition.The specific research contents and results are as follows:1)In view of the problems of the traditional Ada Boost algorithm,such as large number of features,large computation cost,weight overfitting,and high false detection rate in complex cases,this essay suggested an improved algorithm based on eigenvalue center point bidirectional search double threshold.According to the characteristics of face feature distribution curve,the optimal threshold point is found bidirectional parallel with the central point of face feature value as the origin,and the interval range is quickly determined to reduce the time cost.In order to prevent weight overfitting,when updating the sample weight of classification errors,the face sample error rate and non-face sample error rate are introduced to properly adjust the weight of classification errors.By means of cascade of strong classifier,the false detection rate of face in complex background is effectively reduced,which lays a research foundation for face reconstruction and accurate face location recognition.2)Aiming at the shortcomings of current 3D face reconstruction methods such as detail reconstruction,occlusion and wide range of pose,a 3D face reconstruction method based on detail consistency was proposed.The method can generate UV maps from the low dimensional expression space,which is composed of specific face detail parameters and universal face model parameters.The regression device can be trained to predict shape,albedo,expression,pose and illumination parameters from a single image.Using multi-angle loss function to deal with the problem of face reconstruction with large range of attitude deflection,the face image of arbitrary attitude can be reconstructed.The problem of face self-occlusion is solved by symmetric loss function.Finally,the detail consistency loss function is introduced to divide the face details into static details and dynamic details,so as to enrich the details of the reconstructed face,restore the real face in the image as much as possible,and lay the foundation for the subsequent 3D face recognition.3)In order to reduce the impact of the larger parameters of the face reconstruction model on the 3D face recognition speed,the 3D-Mobile Face recognition algorithm was proposed by improving the network architecture on the basis of lightweight backbone network Mobile Vi T and taking into account the speed and accuracy of 3D face recognition.Since it is impossible to predict the influence of face texture features and face structure features on face classification results,a texture-structure attention mechanism is proposed to enable face recognition networks to adjust the weight of texture and structure,increase the weight of features conducive to recognition tasks,and reduce the weight of other features.In order to improve the robustness of the model in the case of face occlusion,a face region attention mechanism is proposed,which makes the model pay more attention to the visible face region and less attention to the occlusion region when face occlusion exists.4)Based on the research of face detection,3D face reconstruction and 3D face recognition,a 3D face reconstruction and recognition system based on front and back end separation is proposed.The system is constructed by Gradio architecture.The H5 visual interface is constructed by Web at the front end,and the 3D face reconstruction and 3D face recognition model are deployed by Flask at the back end,and an API interface is provided for the front end to invoke.Through the performance test of PC,mobile phone and other mobile terminals,the 3D face reconstruction and recognition system has the advantages of cross-platform,easy to use,good interaction and so on.