Robust Face Recognition System in Video using Hybrid Scale Invariant Feature Transform

Abstract

Face recognition plays a significant role in the research field of biometric and computer vision. The important goal of an efficient Face Recognition (FR) system is to have negligible misclassification rate. In video-based face recognition system, the illumination and pose variation problems are predominant. Most of the efficient FR systems are developed for controlled or indoor environment, hence they fail to give accurate recognition in outdoor environment of different illumination variation. Other challenges include occlusion and facial expression. The illumination problem is handled by Histogram Equalization in existing methods. The original Scale Invariant Feature Transform (SIFT) also works well only for pose variation and fails to produce satisfactory results under varying illumination. Hence Hybrid Scale Invariant Feature Transform (HSIFT) with Weighting Factor in feature matching is proposed in this paper which uses a fixed facial landmark localization technique and orientation assignment of SIFT to extract illumination and pose invariant features. The extracted features are then matched using Fast Library for Approximation of Nearest Neighbor (FLANN). The proposed method has been implemented in OpenCV to give a recognition rate of 98% and 95.5% in YouTube celebrity and Extended Yale B dataset respectively.