3D Face Reconstruction From Multi-view Facial Images Based On Morphable Model

In recent years, three dimensional face model is applied to all kinds of fields increasingly, such as health care, education, entertainment, games, animation and so on. Compared to the two dimensional face image, three dimensional face model can reflect more information, including expression,illumination,pose,etcetera.It is because three dimensional face model has a wide application prospect and using value,three dimensional face modeling technique has become a hot research topic in computer graphics, computer vision, human-computer interaction and some other fields,research about this technique has also become more and more deeply.Making use of morphable model to model a three dimensional face has been a new method in recent years. So far, the method used by a lot of people is that making use of single specific face image to model a three dimensional face, this method has only one image, so it lack texture information from different angles and corresponding relationship among different images,itcannot guarantee good results coming from different perspectives each time,Therefore, limitations exist to large extent and the sense of reality is quite low.In order to solve this problem, this dissertation introduces the modeling methods, matching the morphable model of human face with the multi-pictures taken from different perspectives but of same person. In this way, the sense of reality of human face has been greatly improved. The main research contents of this dissertationinclude the following aspects:1. To improve the non-rigid registration method which is based on TPS transformation.We use TPS to represent the non-rigid deformation of three dimensional faces and use FPS to generate the TPS control points automatically.Finally,KD-Tree algorithm is used to reduce time complexity in corresponding points matching search stage,then the algorithm efficiency get improved.2. To dispose the outliers. We propose a method to eliminate the influence of the outliers and we have proved that the method is effective and robust through experiments on BJUT-3D face database.3. Making use of multi-view facial images to model a threedimensionalface.The single view modeling can notguarantee that every angle will produce good effects, so it exists seriouslimitations, in order to avoid these limitations, we have improved the singleview reconstruction algorithm and have finished multi-views reconstruction, it is mainly achieved through improved error function of single view.4.Texture mapping. We project the input two dimensional face image into the three dimensional face model,it will make the final three dimensional face more real. The core concept is to find pixel point and its related information of two dimensional face image in accordance with every vertex of three dimensional face model and let the color of input two dimensional face imageas the final color of pixel, covering the previous pixel.