Symmetric Uncertainty Based Search Space Reduction for Fast Face Recognition

Introduction

Automatic face recognition has been one of the active areas of research in computer vision (Zhao et al., 2003, Narvekar et al., 2008, Phillips et al., 2011, Gou et al., 2014, Du et al., 2014). Human face is the most common cue to identify an individual. Face recognition is one of the most popular and challenging biometric modality compared to fingerprint, iris, gait etc., as it allows unobtrusive identification of people in uncontrolled environment without the need of an individual’s cooperation. Traditionally face recognition has been experimented only on still images. However, the real challenges are only in building a robust face recognition system in video as it has to overcome the variations in illumination and pose, occlusion and facial expression that cause differences in face (Zhou et al., 2003, Patel et al., 2011). Recently, research focus has shifted from image to video-based face recognition (Poh et al., 2010, Choi et al., 2012, Atan et al., 2015). The availability of low cost video cameras and increased processing capability facilitates the process of face recognition in videos at ease. Video inputs are redundant and provide rich information in the form of multiple frames (Zhou & Chellappa, 2005).

Video surveillance systems typically provide continuous video shots of a person at low or medium resolution, from multiple viewpoints and over long period of time. Searching the large database of faces obtained from video frames incur longer time for recognition. Hence, there is a need for reducing the search space in order to have an efficient and accurate face recognition system. The performance of the system depends on how well face patterns are represented, face feature points are selected and recognition models are built for face recognition. Large feature space dramatically reduces the recognition accuracy and exhaustive search mechanism increases the recognition time. The selection of representative feature subsets (Song et al., 2013) is implemented upon discrete random variables but, the process of discretization cannot be used upon image feature points as it is in the continuous space.

The main objective of the study is to develop a methodology that would reduce the search space in terms of reduced features for face representation and reduced templates for face matching, yet achieve acceptable accuracy in face recognition in video. In order to meet the objective, this paper proposes a Symmetric Uncertainty based search space reduction (SUSSR) methodology which involves three stages namely (1) Feature space reduction (2) Search space reduction (3) Classification. In the feature space reduction stage, a subset of useful features that produces similar results as the original feature set is obtained by Symmetric Uncertainty (SU) based graph-theoretic clustering method. In the Search space reduction stage, subsets of features are further partitioned using Fuzzy C-Means (FCM) clustering method. The classification process is considered as a feature matching problem wherein the face from query frame must be compared with a known model that is obtained from a set of images. Kullback-Leibler (KL) divergence (relative entropy) (Kullback, 1997), information theoretic based method is employed which compares the input query face feature with that of the stored facial features for each class from the database. The class with the minimum divergence is reported as the class of the face in query frame. Moreover, in all stages the proposed methodology operates on continuous feature values without a need for discretization. Thus, the proposed methodology reduces the hypothesis space with reduction in the feature space and search space enabling the methods to perform better in terms of efficiency and recognition rate.