Abstract
Face recognition is an important application in the field of pattern recognition. Face recognition technology is commonly used in our daily life, such as e-Gate, access control system, etc. The recognition rate will drop considerably when the head pose or illumination variation is too large, or when there is expression on the face. The greatest challenge is how to get over these difficulties. Face recognition can be divided into the following four steps: face detection, image preprocessing, feature extraction, and feature matching. Histogram of oriented Gradients (HOGs) feature is an effective descriptor for the contour of the face, and robust to illumination effect. So we choose HOG to be the feature in our system. In order to compensate for errors in facial recognition due to rotation changes, we proposed a rotation detection and correction method based on eye-location. Two human face databases, namely FERET database and CMU PIE, are used in our experiments to show the recognition performance of our proposed method and system. In the FERET database experiment, it is found that the rotation degrades the recognition rate drastically, using only HOG feature and subjecting to illumination and facial expression variations. But with the proposed rotation correction preprocessing, the recognition rate of our method is improved over that of the paper [15]. This improvement is also observed in the experiments on CMU PIE database where both EER and FRR/FAR are improved with the proposed rotation correction preprocessing and HOG features under illumination and facial expression variations. Finally, a PC/NB-webcam based face recognition system using the proposed rotation correction preprocessing and HOG features and the combined FERET and PIE face databases is implemented. Some experiments show the rotation and illumination resilient property of our proposed system.