| Automatic recognition of human faces is an important and growing field. Several real-world applications have started to rely on the accuracy of computer-based face recognition systems for their own performance in terms of efficiency, safety and reliability. Many algorithms have already been established in terms of frontal face recognition, where the person to be recognized is looking directly at the camera. More recently, methods for non-frontal face recognition have been proposed. These include work related to 3D rigid face models, component-based 3D morphable models, eigenfaces and elastic bunched graph matching (EBGM).;This thesis extends recognition algorithm based on EBGM to establish better face recognition across pose variation. Facial features are localized using active shape models and face recognition is based on elastic bunch graph matching. Recognition is performed by comparing feature descriptors based on Gabor wavelets for various orientations and scales, called jets. Two novel recognition schemes, feature weighting and jet-mapping, are proposed for improved performance of the base scheme, and a combination of the two schemes is considered as a further enhancement. The improvements in performance have been evaluated by studying recognition rates on an existing database and comparing the results with the base recognition scheme over which the schemes have been developed. Improvement of up to 20% has been observed for face pose variation as large as 45°. |
