Scale and pose invariant face recognition

Face recognition is an active research area which is quite attractive nowadays and has commercial and law-enforcement applications have been deepened and become widespread day by day. Through the advances in imaging technology and its widespread usage, there exist many challenging issues seen on 2D face images especially on which have been taken in real world conditions. These challenges are basically based on scale, pose, illumination, face expression variations, occlusion and aging. In this thesis work, a face recognition system which is robust against especially scale and pose challenges and also executable as real-time and full-automatic has been proposed. For face detection, which is first step of face recognition, conventional Viola-Jones method has been incorporated. After face detection step, in order to cope with challenges due to scale and pose variations, which are fundamental concerns of this thesis work, face alignment step has been added. In order to align face images, Constrained Local Models which is remarkable with its success in last years and executable as real-time, has been used. Features have been extracted from aligned face images as component-based and holistic based. In order to extract features, Gabor Wavelets and Local Binary Patterns (LBP) have been employed. Extracted features has been classified by linear Support Vector Machines directly or in a dimension reduced by Principle Component Analysis way. Success of classification has been evaluated by 5-fold cross validation.xiiiThe proposed method has been applied on IMM, AR, FG-NET, Caltech and MUCT face databases. Results of experimental works have been compared with regards to feature extraction methods, which are Gabor Wavelets or LBP, general approaches are component-based or holistic-based and whether dimension is reduced by PCA.According to the results obtained from experimental works, component-based approach, extracted features via Gabor Wavelets and dimension reduced by PCA produces more accuracy rates generally in comparison to other configurations. In addition, when the result has been also compared to current works in literature, success of the proposed method can be seen.Finally, it has been concluded that the proposed method fits its purpose very well with the high accuracy rates on 2D face images which are acquired easily from real world with scale and pose variations.Key words: Face Recognition, Constrained Local Models, Gabor Waveletes, Local Binary Patterns, Support Vector Machines