Facial Feature Automatic Location And Facial Recognition Based On Improved Active Appearance Models

As one of biometric technology, facial recognition is based on the known face database, using image processing and pattern recognition technology to identify or verify one or more personal faces from a static or dynamic scene, and having broad application prospects in commercial, security and other fields. Facial feature location is the key to facial image analysis, and the studies on facial recognition and facial feature location have important theoretical and application value. Facial feature location is automatically positioning various features at the human face image by computer, including the position of eyes, nose, mouth, chin, ears and the profile of the human face. Active Appearance Models (AAMs) is an effective method of facial recognition and facial feature location, and has become a research hotspot in image processing areas. This thesis is mainly study the facial feature location and human face recognition based on Active Appearance Models. The main job and innovation as follows:1. This thesis presents the basic idea and method of Active Appearance Models in detail, puts forward an improvement AAMs based on Independent Component Analysis (ICA), and uses the ICA-AAMs in facial feature location and facial recognition experiments. The classic AAMs adopts Principal Component Analysis (PCA) as the statistical analysis. PCA only takes into account the second order statistical information, and is difficult to extract the local features. So the classic AAMs is poor at describing the local features of the human face. The ICA is based on higher statistical information, and the components yielded by ICA are not only not correlative but also independent. So ICA has a better performance in describing the local features of image. PCA results in a natural ordering of the Principal Components according to their variance, while with ICA, such an ordering is not obtained automatically. This thesis proposes a method for ordering which to order independent components according to the variance between their modules and the mean module. In the ICA-AAMs, we adopt ICA as the statistical analysis in statistical shape model and texture model.2. A novel method for building 3D human face model is proposed. It makes use of the depth information and corresponding intensity information generated by Correlation Image Sensor (CIS), and extends the 2D AAMs to improved AAMs which is based on facial depth and intensity information. The classical 2D AAMs combines the shape and texture information of the object, and is an effective method to build 2D deformable model. But it can't describe the object in 3D space. This thesis makes use of the facial depth map information and its corresponding intensity image to build face depth model, and fuses the face shape model, depth model and texture model to build the improved face active appearance model. Meanwhile, the target function of AAMs matching is modified to the weighted sum of depth error and texture error. And the weighted coefficient is determined through experiments.3. The initial pose parameter is given out by key points of the facial feature using curvature of 3D facial surface. Because the initial pose parameter is an important factor which affects the accuracy of AAMs, we use the curvature of 3D facial surface to complete preliminary location of the facial features, and take the location result as the initial pose parameter of the improved AAMs matching. This thesis discusses the process of the improved AAMs and its matching, the calculation of curvature of the facial surface and the preliminary facial feature location algorithm based on curvature.