Face recognition in video using deformable parts model with scale invariant feature transform (DPSIFT)

Abstract

Face recognition is a complex task due to the challenges of varying pose, illumination, scaling, rotation, and occlusion in live video feed. This paper proposes a hybrid approach for face recognition in video called Deformable Parts Model with Scale Invariant Feature Transform (DPSIFT), to make face recognition system invariant to illumination, scaling, rotation, and limited pose. The proposed method identifies the significant points of the face using deformable part model and SIFT feature descriptors are extracted for those significant points. Fast Approximate Nearest Neighbor (FLANN) algorithm is used to match the SIFT descriptors between gallery image and probe image to recognize the face. The proposed method is tested with video datasets like YouTube celebrities, FJU, and MIT-India. DPSIFT method was found to perform better than the existing methods.