Research On Face Recognition With Pose Variation

Face Recognition, which has been studied for more than 30 years, is one of the challenging researches in pattern recognition and machine vision, it can be widely applied in the public security, information security and human-computer interaction. By now, face recognition technology under well-controlled imaging condition is practically usable, while the performance of algorithm dramatically decreases under uncontrolled environment, which involves the variation of head poses and illumination et al. Robust face recognition is one of the key issues. Face recognition with pose variation are studied in this thesis.MR-ASM(Multi-resolution Active Shape Model)and AAM(Active Appearance Model)are widely used in feature extraction of face recognition. The most important advantages of AAM and ASM are the capabilities of global restraining and robustness to obstacle. In this thesis, two improved ASM and AAM algorithms are proposed and the pose-irrelevant features are extracted for face recognition:1. Hybrid MR-ASM and AAM for location of facial landmark points.MR-ASM and AAM are combined flexiblely in this algorithm. AAM is initiated using the final result of MR-ASM. In the image of the highest resolution, ASM and AAM are used iteratively. And the location of landmark points is guided using both shape and texture. The contribution of this algorithm involves: 1) The procedure of computing controlling parameters of AAM is improved, so that the parameter for both shape and texture could be obtained simutanuously. 2) The parameters update method is improved in the location of MR-ASM, in order to take the mutual effection of variation of shape and pose into consideration.2. Cascade-MR-ASM for location of facial landmark points.In this algorithm, the whole training sample set is partitioned to several subsets automatically based on the objective: making location of landmark points more accurate. All the independent MR-ASM locators trained from corresponding subsets are organized in cascade for location of facial landmark points. By the objective, the distribution of shape control parameter in each subset is more clustered. It also avoids the problem of estimating the distribution of parameters.3. Face recognition using pose-irrelevant features.Face recognition is accomplished by the result of landmark point location using the above two algorithms. First the public parameter space is established using the groundtruth location of training set. The locations of landmark points obtained by the above algorithms are mapped into the public parameter space, and the corresponding shape control parameters are obtained. Then the pose-irrelevant parameters are extracted and input to the Near Neighbour Classifier and BP Neural Network classifier respectively for face recognition. Finally, the influence of pose relenvant parameters is analyzed.In this thesis, part of CMU-PIE face image database is used to test our algorithms. The experimental results show that both proposed algorithms can locate the facial landmark points effectively for faces with pose variation. Furthermore, the performance of face recognition based on Near Neighbour Classifier and BP Neural Network Classifier are improved by using the pose irrelevant features extracted from the output of the above two algorithms.