Pose-tolerant Non-frontal Face Recognition using EBGM

Traditional automatic face recognition methods focus on handling frontal or near frontal face images. Therefore, they cannot be directly applied to the pose-varied or non-frontal face images captured by non-intrusive video surveillance systems. In this paper, a non-frontal face recognition algorithm based on elastic bunch graph matching (EBGM) is proposed. The proposed method measures face similarity using facial features which are more robust to pose variation. Experimental results show that the proposed method can achieve a verification accuracy of 97% on face images with 30deg pan-angle. Also, the proposed method can reasonably tolerate plusmn10deg variation in the pan-angle, indicating its robustness in tolerating errors in pose estimation. This method can be easily extended to provide non-frontal face recognition up to half profile.