The Research Of Pose Variation Face Recognition And The Key Technology

Because of the potential application value in personal identification and video surveillance, face recognition technology has always been one of the hot spots of pattern recognition field. The existing technologies perform really well under ideal conditions, but however the vast majority of face recognition perform badly on the occasion of pose and light conditions change. Pose problem, in particular, has become one of the bottlenecks restricting the development of face recognition technology.In allusion to the problem that the change of face pose will cause serious interference to the recognition result and the status that at present most face recognition systems only support standard frontal face, this paper put forward a method which can deal with non-frontal face images based on using Gaussian process regression to analysis the relationship between frontal and non-frontal face. The core idea of this method is to make pose correction to the input face images before face identification. In this paper, we do the following work on the purpose of pose correction of non-frontal face images:1, In order to achieve the purpose of automatic recognition, we have trained face and face landmark point detector based on Adaboost and random forest algorithm. These detector can detect face and face landmark points automatically.2, We have trained multi-pose active appearance models, then we can find the most matching model by pose estimation or face landmarks and accurately extract the face contour of input images, implement automatic face under posture change environment.3, Study a variety of prediction strategies to fitting the frontal face contour, and finally determine use nonlinear strategy which is represented by Gaussian process regression. We have trained several frontal face contour prediction models, and through the piecewise affine transformation to wrap the face texture feature to the predicted frontal face contour. By this way, frontal face images can be synthesized. Recognition experiments were conducted on the Multi-Pie and FERET databases, showing high accuracy across these databases.