3D Human Face Modeling And Its Application Based On Images

The creation of realistic 3D virtual faces has always been considered as the Holy Grail of computer graphics.Human face has a complex structure,and contains varied facial expression movements.Realistic face modeling requires the restoration of facial features and details as fully as possible.With the development of technology,large-scale hardware scanning equipment has been able to capture the details of the face,but at the same time the demand for small and simple face modeling system is growing,and people hope that face modeling technology can get rid of the complex scanning devices.Since image is the most common form of face information,reconstructing 3D face model from image becomes the main direction of the research.This dissertation focuses on several aspects of 3D face modeling based on image,and proposes several novel methods for reconstructing low resolution 3D face model and high resolution 3D face model from images,and establishes a complete image face transfer system.The main work and innovation of this dissertation are as follows:1.The low resolution face model reconstruction algorithm based on single image is studied,and an novel method combining face statistical model and surface Laplacian deformation is proposed.Reconstructing 3D face from a single image is an under constrained problem,which requires to be solved by a priori knowledge of face statistics.In order to eliminating the reconstruction error introduced by the face statistics,the reconstruction results of face statistics model are modified by Laplacian deformation.The local surface of the face model is deformed by Laplacian deformation,which further improves the reconstruction precision.The experimental results show that the 3D face model produced by this method has better overall effect,and the reconstruction precision near the model feature points is high,which is consistent with the result of the 2D feature point detection.2.A novel 3D face reconstruction method based on random image collection is proposed.Random image collection is a set of multiple photos of the same face in a variety of occasions under random shootings,common in personal album or web search.In order to eliminating the ambiguity solution of the 3D face reconstruction from a single image,the single image is expanded to a random image collection,and the 3D face model is reconstructed by combining multiple random images.Since the random image collection has the same face identity feature,by jointly solving the low resolution model from multiple images using face statistical model,the ambiguity can be effectively reduced.The experimental results show that the 3D face model reconstructed from multiple images is more close to the target face in the image,and the effect is better than the reconstruction result based on single image.3.The face detail recovery algorithm based on Shape from Shading(SFS)is studied,and an improved method combining the illumination constraint and the image gradient constraint is proposed.On the basis of obtaining the low resolution model of face,the geometric details of face surface are restored from the shading information to enhance the reality of face model.Aiming at the problem that the traditional SFS method is not robust to the illumination condition,an improved method is proposed.By introducing the new constraint condition of the illumination and face surface normal,the stability of the optimization is improved.The experimental results show that the method can effectively deal with the problem of face detail restoration under complex lighting conditions.The reconstructed effect is better than that of the traditional SFS method.4.A joint high-resolution model reconstruction algorithm based on random image collection is proposed.SFS method can not effectively deal with the image region where high light or self-occlusion shadow occus,consequently pseudo-details will be created in these regions.At the same time,the result of SFS is the scattered point cloud,which needs to be further rebuilt into a high resolution model.Aiming at the above problems,a novel high-resolution model reconstruction algorithm is proposed.By reconstructing the low resolution face model from the random image collection,the face detail is further restored,and the depth point cloud is obtained.Then,the high-resolution models of random image collection are solved by combining the surface deformation transfer algorithm.Experiments show that this method can eliminate the pseudo-detail to a certain extent.5.The image face transfer technology is studied,and a complete image face transfer system is established.Image face transfer is an extended application for low-resolution reconstruction of human face.The technique translates the user's face image into the target image,replaces the face in the target image,and naturally fuses the user's face into the target scene.In order to match the illumination condition,color distribution and image contrast between the user image and the target image,an image fusion algorithm based on Laplacian pyramid is proposed,which can adaptively adjust the color distribution and detail contrast of the image.In order to solve the problem of partial occlusion in the target image,an image fusion algorithm for partial occlusion is proposed,which can achieve a smooth transition of occluded objects edge to face,and can restore the cast shadows on the face area.Based on the above technologies,a complete image face transfer system is established,and the face transfer test is carried out under the condition of no occlusion and partial occlusion respectively.The experiment shows that the system can get better effect,better than the traditional methods.