New face recognition descriptor based on edge information for surgically-altered faces in uncontrolled environment

Abstract

Since plastic surgery have increasingly become common in today’s society, existing face recognition systems have to deal with its effect on the features that characterizes a person’s facial identity. Its consequences on face recognition task are that the face images of an individual can turn out to be distinct and may tend towards resembling a different individual. Current research efforts mostly employ the intensity or texture based descriptors. However, with changes in skin-texture as a result of plastic surgery, the intensity or texture based descriptors may prove deficient since they enhance the texture differences between the pre-surgery and post-surgery images of the same individual. In this thesis, the effect of plastic surgery on facial features is modelled using affine operators. On the basis of the near-shape preserving property of the combination of the operators, the following assumption is made: The edge information is minimally influenced by plastic surgery. In order to exploit this information in real-world scenarios, it requires that face images be evenly illuminated. However, an evenly illuminated face image is far from reality on applying existing illumination normalization techniques. Thus, a new illumination normalization technique termed the rgb-Gamma Encoding (rgbGE) is proposed in this thesis. The rgbGE uses a fusion process to combine colour normalization and gamma correction, which are independently adapted to the face image from a new perspective. Subsequently, a new descriptor, namely the Local Edge Gradient Gabor Magnitude (LEGGM), is proposed. The LEGGM descriptor exploits the edge information to obtain intrinsic structural patterns of the face, which are ordinarily hidden in the original face pattern. These patterns are further embedded in the face pattern to obtain the complete face structural information. Then, Gabor encoding process is performed in order to accentuate the discriminative information of the complete face structural pattern. The resulting information is then learned using subspace learning models for effective representation of faces. Extensive experimental analysis of the designed face recognition method in terms of robustness and efficiency is presented with the aid of publicly available plastic surgery data set and other data sets of different cases of facial variation. The recognition performances of the designed face recognition method on the data sets show competitive and superior results over contemporary methods. Using a heterogeneous data set that typifies a real-world scenario, robustness against many cases of face variation is also shown with recognition performances above 90%.