Complex Context Of Multi-pose Face Recognition Technology Research

Facial representation and recognition is one of the key issues in pattern recognition,computer vision,and image understanding fields. It is widely applied to numerous commercial and law areas,such as mug shots retrieval,real-time video surveillance,security system and so on. Now recent researches are focused on face recognition with frontal or frontal plus profile in a simple background,and multi-pose face recognition from an image with a complex background are just beginning.The thesis aims to study the fundamental theory and key technology of multi-pose face recognition from an image with a complex background. The issues of multi-pose facial detection,organic location,feature extraction and recognition in an image with a complex background are discussed in details. It includes the following research works.In face detection in an image with a complex background,the methods of face detection in gray images are summarized and analyzed. Then an algorithm of face detection in color images is developed based on skin-color and multi-level verification. Based on the statistical analysis of skin-color,the candidate face regions by color segmentation are detected. Then shape analysis,facial local feature,head model are applied to verify very candidate region to get real face regions. Applying it to more than 120 color images,experiment results show that the multi-pose face with two visible eyes in color images can be detected.In facial feature extraction,it includes three aspects.(1)A multi-pose face image normalization method is presented based on two pupils and the center point of the lip. It can save the information near lips,which may be deleted by the method based on two points.(2)Based on the analysis of current methods,a new multi-pose facial feature location algorithm is developed,which is based on the analysis of multi-feature and integral projection,the combination of an iterative search with a confidence function and template matching. The algorithm not only improves the location accuracy,but also speeds up a great deal.(3)Based on the analysis of the advantages and disadvantages of current feature extraction methods,an adaptive facial feature selection criterion is developed,which is based on facial local feature protrusion consisting of several aspects,such as face image resolution and image quality. Experiment results show stable and efficient facial features can be selected.In multi-pose face recognition based on multi-views,the concepts and methods of traditional frontal and multi-pose face recognition are firstly summarized. Then a multi-pose face recognition algorithm is presented based on a hierarchical model with a pose subdivision and fusion decision. It first does face recognition in every sub-pose according to the hierarchical model and then gets the final face recognition results by a fusion decision algorithm. The algorithm not only enhances the accuracy,but also reduces the computation.In multi-pose face recognition based on a single view:(1)the technology approaches of multi-pose face recognition are firstly studies based on single view or small samples in details.(2)A new multi-pose face recognition algorithm is developed based on single view. It first generates the multi-pose face images based on a fit method with a high order polynomial function. Then,it does face recognition based on the single view and the generated multi-pose images. Experiment results show its performance exceeds the traditional methods a great deal.(3)Several multi-pose face image synthesis methods are in details studied,which is the key technology of multi-pose face recognition based on a single view.Based on the above research work,a multi-pose face recognition technology in images with a complex background is proposed and then the model system is implemented. The model system can accomplish color multi-pose face detection,organic location and feature extraction,and recognition from static images with a complex background and two visible eyes. Its experiment results show its robustness and potentialities.