Parametric Model Of 3D Face Reconstruction And Face Model Rendering

With the improvement of computing power and the development of 3D acquisition methods,face reconstruction has entered a new era.Face reconstruction can be applied to multiple scenes such as movies or animation,games,social and shopping.The development of face reconstruction technology can greatly promote the development of 3D industry,and enhance the user experience in related products.In this thesis,facial feature points are treated differently according to their locating features.We develop a method for 3D face reconstruction using a novel landmark updating optimization strategy.Our method is also based on minimization of the landmark projection error for calculation speed.To achieve more accuracy,contour landmarks and self-occluded landmarks need to be updated.For contour landmarks,the detected landmarks on 2D image are to updated by finding a nearest position on contour curve.For self-occluded landmarks,we render the model onto image plane,extract the edge of projected area,and generate new correspondence landmarks according to the edge pixels.Finally,Laplacian-Beltrami operator is applied to face model for facial expression.We test our method quantitatively with ground truth data,and provide example reconstructions.The experiments result shows that our method performs better on MICC dataset.After the face model reconstruction mentioned above,we analyze the rendering process of the Lambertian surface reflection model,and gives a derivation of the reverse rendering process based on the Retinex hypothesis etc.So that we can extract more detailed information from the image.We propose two methods to add details:to directly modify the original model coordinates or to use normal maps.By comparing the advantages and disadvantages,we finally use the normal map to give the face model more details.Normal maps can overcome the fact that the model itself can't show more detail,so this method can make full use of the details contained in the input image and give a detailed visual effect when rendering.By merging the face model into a full head model,we can obtain a more realistic model.This step uses the Laplacian operator to keep the normal direction of the seam smooth,making the merged seams more natural.In addition,the hair represented by mesh patch can be made for the human head model and converted into a hair model for rendering.Through the above post-processing,the realism of the model can be greatly increased.