Face Recognition Algorithm Based On Feature Fusion

Face recognition as an application field of pattern recognition, computer vision and machine learning has great theory value, and has been widely used in many commercial fields such as security, human-computer interaction, and video surveillance. At present, under controlled conditions, face recognition technology can obtain better results. But when conditions are uncontrolled, such as changes of illumination and pose, the recognition result will decline sharply. Therefore, it is a great challenge for face recognition research as illumination and pose vary.Feature extraction is the core issue of face recognition. The influence of illumination on face recognition is researched and robust illumination face feature representation is designed in this thesis. Multi-scale LBP histogram statistics can extract structural information at different levels and better characterize global features of the face, and face representation based on Gabor wavelet is the most successful representation of local features. With Local Binary Pattern(LBP) and Gabor wavelet have robust illumination characteristics, a fusion of multi-scale LBP histogram features and fast Gabor features of regional cascade is proposed. To extract more robust illumination characteristics, before feature extraction, wavelet reconstruction is used to normalize the illumination in face images.A novel fusion rule based on the principle of Dempster-Shafer evidence theory is proposed to fuse global and local features. Dempster-Shafer evidence theory is the extension of classical probability theory and provides a powerful method for the expression and synthesis of uncertainty information, especially for fusing decision information. Using the personality of a specific person, a face verification method based on feature reconstruction of personality principal component analysis is used.A robust prototype system of face recognition is implemented under variant conditions. The test results on the large-scale face database CAS-PEAL-R1 show that our algorithm can achieve better recognition performance under variant illumination and different expressions.