Face Recognition Using Multi-local Descriptors—A Novel Approach

Abstract

While face recognition has been a topic of interest for the researchers for quite some time now, most of the advancements and superior results have come out from the field of face recognition in controlled scenarios. As we shift away from the controlled environment like passport images, driver’s license etc. to unconstrained environment like images taken from surveillance footage, images taken by bystanders etc. the recognition accuracy significantly decreases. This inconsistency is due to the fact that face images in unconstrained environment have vast variations in parameters like illumination, background detail, pose, expression, occlusion etc. At present times when important disciplines like security and forensics application depends on such systems it can prove to be very useful if a face recognition system in unconstrained environment can give comparable results to the systems in controlled environment. Our work focuses on face recognition using deep feature extraction by concatenating the features of different feature extractors to improve the recognition accuracy in unconstrained environment. We use multiple feature based methods (variants of LBP and LGS) to extract important features from the same image and combine them to form a single feature vector. For classification SVM is used and two face databases viz. ORL face database and LFWCrop (Labeled Faces in the Wild). The experimental results reveal that the proposed method improves the performance of the face recognition system.