AAM Human Face Fitting Based On Texture Weight

Human face fitting is an important topic in computer vision, which has been widely applied to virtual reality, human face pose estimation, intelligent control, face recognition and other fields. Human face technology has gone into the daily life. Because the face is three-dimensional space objects, which's two-dimensional projection has a large amount of data. And because of the varied facial expressions, it is difficult to model human face accurately.In addition, some outside interference, such as uneven lighting, shelter materials, etc., adds to the difficulty of fitting the face. Over the years it has been conducted in-depth study on this issue, resulting in a lot of face modeling methods. However, most treatments of test images have stringent requirements, such as illumination conditions, the front face. And to real time or the accuracy,it still can't meet the actual requirements. Our method is one of the most widely used methods, namely, the AAM human face fitting method. And add an concept of texture weight on this basis. Our method can make sure it has the same accuracy with the traditional AAM method. At the same time,it reduces the computing time greatly and improves the efficiency.The details of our studies as followingsExtract image feature points in the training set, then apply linear PCA dimension reduction to it. Then we get shape feature vectors Si, combine which with the shape parameters P, we can constitute a shape model. This kind of face shape modeling method can reduce the calculation complexity, will accelerate the computing speed greatly and improve the modeling efficiency.As the shape modeling, we use the raster scan to extract the pixel values of the training set and stored as textures. Then we normalize the textures, after that apply PCA linear dimension reduction to it, we get the appearance feature vectors, Ai, combine with the given parametersλi we can build the appearance model. Same as shape modeling, it can improve the modeling efficiency.After select the test face image, according to the initial value of parameters P which the initial face detector give, we can build the initial shape S. Then we warp A(x) defined in the interior of the base mesh so to current shape S with piecewise affine warp PAW,so we get an AAM model instantiation.To fit the test image with the model instantiation. We use the image fitting strategy "fitâ†'compareâ†'adjustâ†'re-fit and re-compareâ†'re-adjust", adjust the shape parameters P and the appearance parametersλi constantly, then we fit the model to the test image exactly finally. The final positions of the shape feature points is the facial features to match.Traditional AAM fitting treats facial feature points and non-feature points equally. But in fact the role of feature points is more important than non-feature points in the fitting process. The AAM face fitting base on the texture weight in this paper,that is, we multiply AAM appearance model with a texture weight function, this function's value according to the distance between the pixels and the feature points. The feature points closer to the feature points get the greater weight. It highlight the role of the feature points in the fitting process, and relatively weaken the role of non-feature points, thus we can reduce the adverse effects of the non-feature points.