3d Face Reconstruction Based On A Single Image And Statistical Model

Face has a very complex physical structure, which is rich in important physiological and psychological information, and it is also a very important and complex communication channel. Face modeling and facial animation is one of the most fundamental and most difficult issues in computer graphics and computer vision fields, while the problem is also a cross-disciplinary issue of physiology, psychology, physics and other fields. Study and resolution of the problem can provide a reference for complex modeling and offer important promotion for similar scientific issues.In addition to important scientific significance, realistic face modeling also has a wide range of practical applications. Face modeling and facial animation have broad application prospects in many fields, such as film and television production, video telephony, video conferencing, computer games, face recognition, human-computer interaction, virtual reality, medical science, public safety, etc.For now, image-based 3D face reconstruction technique a popular topic both in international and domestic research. Because the technique of 3D face reconstruction based on a single is more difficult, and there is little relevant content in our country, this thesis provides a profound research on the 3D face reconstruction methods based on a single image and statistical model, and propose a new method both for building a statistical model and for efficient face reconstruction. Specific study points are shown as follows:1) Based on the In-depth analysis and summary of previous achievements, we discuss current researches and applications of 3D facial reconstruction technology in detail and mainly focus on a single-image based face reconstruction method.2) Study general method of the establishment of three-dimensional face of the statistical model and proposed a new statistical modeling approach.The core step of building three-dimensional face statistical model is to establish a point-to-point correspondence among 3D faces in database. As the data points of 3D faces are huge in size and face has a complex physiological structure, thus establishing point-to-point correspondence is a very difficult problem. We use thin plate splines to establish the dense point-to-point correspondence by combining two-dimensional and three-dimensional information.3) Propose an efficient face reconstruction method based on a single image.Face reconstruction only uses a frontal face image. The reconstruction process consists of three-dimensional shape reconstruction and texture reconstruction. Based on the landmarks, we use fitting algorithms to optimize shape coefficients, thus the corresponding three-dimensional face shape can be reconstructed. In order to improve efficiency, we use texture mapping to reconstruct texture, which can guarantee the realistic reconstruction while dramatically reduce the reconstruction time.For in-depth studies of above aspects, this thesis has made some progress and performance in the following areas:1) Put forward a new method to establish the dense point-to-point correspondence among 3D faces by combining with two-dimensional and three-dimensional information, and build a statistical model of BJUT-3D face database;2) Reconstruction based on only a single frontal image;3) The three-dimensional reconstruction time has greatly reduced, only 4s implemented by MATLAB...