Research On Reconstruction Of Face 3D Model Based On Single Image

3D face reconstruction is an active research topic in the field of computer vision.It has a wide range of application scenarios,such as character design in game animation,interactive applications in the field of virtual reality and simulation modeling in medical field,etc.Although 3D face reconstruction has been studied for decades,3D face reconstruction is still a very challenging research topic due to the diversity and complexity of faces,especially to complete 3D face reconstruction with a single face image.Compared with 3D face reconstruction based on multi-views,3D face reconstruction based on a single image requires less input information,which enables us to complete 3D face reconstruction more quickly.Compared with the expensive 3D reconstruction equipment to restore 3D face structure,3D face reconstruction based on a single image is cheaper and easier to operate.Therefore,the research of the subject is of great value.This paper improves the single-face 3D reconstruction algorithm based on deep learning through encoding-decoding architecture.In addition,a hierarchical feature fusion module and an asymmetric loss function based on weight mask are proposed to improve the accuracy of 3D face reconstruction.Finally,a head pose estimation system based on 3D face reconstruction is designed.So the work of this paper mainly includes:Firstly,this paper improves the Pr Net(Position Map Regression Network)3D face reconstruction algorithm based on deep learning method,and proposes an Efficient Net hierarchical fusion codec network based on attention model.In the encoding network,the lightweight Efficient Net is used as the backbone network,and a hierarchical feature fusion module is proposed.Moreover,the attention model mechanism is introduced into the decoding network,which makes the whole network model pay more attention to face features and improves the accuracy of 3D face reconstruction.Then,in order to solve the problem that the mean square error loss function treats all points equally,an asymmetric weighted mask mean square error loss function is proposed.This loss function can make the network pay more attention to learningfeatures of the face.And the contribution to the loss function is different for the features of different regions of the face.Finally,this paper proposed a head pose estimation system based on 3D face reconstruction.The system continuously loads each frame of data in the camera,then uses the Dlib library to complete face detection,and uses the Pytorch framework to complete the algorithm construction of the Efficient Net hierarchical fusion codecs network based on the attention model.And input the detected face into the algorithm to complete 3D face reconstruction,and then calculate the head pose parameters according to the reconstruction results.Finally,a WEB service was built by Flask framework to display the results of 3D face reconstruction and head pose estimation in real time on the WEB page.Experimental results show that compared with other 3D face reconstruction methods,the proposed algorithm can quickly reconstruct 3D face model.In the GPU environment,it takes only 5.3ms to get the result of 3D face reconstruction.At the same time,the accuracy of 3D face reconstruction is kept at an acceptable level.The NME(Normalised Mean Error)of the proposed algorithm for 3D face reconstruction is 4.09%.Compared with Pr Net 3D face reconstruction algorithm,the performance of the proposed algorithm loses about 3% in reconstruction accuracy,but the reconstruction speed is improved by 45%.