Research On The Technology Of 3D Face Reconstruction Based On Deep Learning

In the field of maxillofacial plastic surgery,doctors need to model the face of a patient in order to develop a surgery or treatment plan.This thesis studies the 3D face reconstruction technology based on deep learning,and realizes a set of 3D face model imaging,reconstruction and measurement analysis system for clinical application of maxillofacial plastic surgery.Stereo matching is the key step in 3D surface reconstruction technology.Effective image feature extraction is an important basis for stereo matching.This thesis proposes a stereo matching algorithm combining context contrasted features and multi-scale feature fusion.The adoption of context contrasted features can solve the problem of unbalanced feature information of significant targets and non-significant targets in images.Multiscale feature fusion can effectively extract different scale feature information of the same target in images.In this thesis,the experimental results show that the parallax’s 3-pixel matching error rate can be reduced by 4.57 %,which effectively improves the matching accuracy of the stereo matching algorithm.The ill-conditioned problems such as occlusion,low texture and inconsistent light in the original images are important problems for the stereo matching algorithm.This paper proposes a stereo matching algorithm based on the encoder-decoder architecture.A spatial pyramid network module that can extract multi-scale feature information and a 3D convolution network that can extract global context information are added to the network structure of the algorithm.Parallax regression uses a differentiable soft argmin function to directly return sub-pixel disparity values from the parallax cost.Experimental results show that the algorithm effectively reduces the mismatch probability caused by the illconditioned region,and the matching speed can reach 100 milliseconds/frame.The face configuration is unique,covering a range of nearly 200 degrees from the left ear to the right ear.It is difficult to obtain sufficiently complete facial information by simply acquiring images using a binocular imaging system.In this thesis,an image acquisition system based on multi-view stereo vision imaging principle is designed,which effectively expands the imaging field of the imaging system.The acquisition system integrates two binocular imaging systems to acquire images of the left and right faces,respectively,providing a complete facial image for the reconstruction algorithm.The system implemented in this paper based on this program has been applied in the clinic.Based on the above works,this thesis implements a real face 3D reconstruction and measurement analysis system based on deep learning.The system adopts the face reconstruction technology of passive stereoscopic vision,which can realize the goal of rebuilding high-quality realistic face model in one shot.At the same time,in order to meet the needs of clinical application,the system also realizes the basic measurement function of the distance,area,curvature and other parameters of the 3D face model and the registration function of the double face model,which can help doctors to measure and analyze the preoperative and postoperative effects.