Research On Fitting Algorithm Of Face Construction Base On Deformable Model

Three-dimensional reconstruction is an important topic in the field of computer vision, especially 3D face reconstruction, because it has a wide range of applications in games, animations and movies. There are lots of methods in face construction, the traditional method include: profile-based, color information-based, texture-based. Some methods can reach a high construction precision, but they need calibrate camera, however, some methods do not require calibration but with low precision, the face construction method which based on the deformation model is a good solution to this problem.A new fitting system is built based on a morphable model, in our system several cost functions were used, so the overall function become smoother. To solve this optimization problem we use an efficient algorithm called stochastic newton method. During fitting several strategies are also applied. Compared with the traditional methods our algorithm surpass in accuracy and stability.The main framework of our algorithm is an iterative process, it has two phases: image analysis and image synthesis. In the image synthesis stage the deformation model are rendered to an image. In the image analysis stage we develop several cost functions, and us stochastic newton method to solve. Firstly, building a deformable model. Then, we apply image rendering step with the knowledge of computer graphic, do geometric transformation, rasterization and phong shading by writing shader program. Here, a depth buffer technic was used to make a shadow map and the final rendered image look more realistic with it. Secondly, followed by the image analysis stage, here a feature cost function, an outline cost function and a pixel intensity cost function between the input image and the rendered image are used. Feature points are fitted to adjust model's pose, contours are fitted to adjust model shape, image pixels are fitted to adjust model shape and texture, in a word to make the difference between input image and rendered image smaller and smaller. During fitting stage several strategies are used to make our algorithm more robust. The output of previous fitting stage are as input of next stage, so it is close enough to the true value and can be solve stability. In the last, the time complexity of entire fitting system has been further optimized.two experiments were complished between our system and the business face construction software called Face Gen. The first set of experiments show that our algorithm have more advantageous in accuracy. The second set of experiments prove that our algorithm is more robust with little restriction.