Recognizing Face Images with Disguise Variations

Automatic face recognition, required in law enforcement applications such as surveillance, border security and forensic investigation, is a process in which an individual is identified or verified based on facial characteristics. Researchers have proposed several algorithms that can effectively recognize individuals in controlled environment with minor variations in pose, expression, and illumination (Zhao et al., 2003), (Li & Jain, 2005), (Wechsler, 2006), (Delac & Grgic, 2007). In recent face recognition test reports such as FRVT 2002, FRGC 2004, and FRGC 2006 (Philips et al., 2006 & Philips et al., 2007), the results show that under normal changes in constrained environment, the performance of existing face recognition systems is greatly enhanced. However, in most real world applications, images may not be of good quality or user may not be cooperative or there may be temporal variations and dissimilarities in facial characteristics that are artificially created using disguise accessories. Challenges in automatic face recognition can be classified into six categories: illumination, image quality, expression, pose, aging, and disguise. Among these challenges, recognition of faces with disguise is a major challenge and has only been recently addressed by few researchers (Alexander & Smith, 2003), (Ramanathan et al, 2004), (Silva & Rosa, 2003), (Singh et al., 2008). As shown in Fig. 1, the inter-personal and intra-personal characteristics can be modeled using disguise accessories to alter the appearance of an individual, to impersonate another person, or to hide one’s identity. For example, a criminal can alter facial features and appearance using makeup tools and accessories to remain elusive from law enforcement. The challenges due to disguise cause change in visual perception, alter actual data, make pertinent facial information disappear, mask features to varying degrees, or introduce extraneous artifacts in the face image. Existing face recognition algorithms may not be able to provide the desired level of security for such cases. In literature, Ramanathan et al. (Ramanathan et al., 2004) studied facial similarity for several variations including disguise by forming two eigenspaces from two halves of the face, one using the left half and other using the right half. From the test image, optimally illuminated half face is chosen and is projected into the eigenspace. This algorithm has been tested on the AR face database (Martinez & Benavente, 1998) and the National Geographic database (Ramanathan et al., 2004) which consists of variations in smile, glasses, and illumination. An accuracy of around 39% for best two matches is reported on the AR database. Silva and Rosa O pe n A cc es s D at ab as e w w w .in te ch w eb .o rg