Elastic Bunch Graph Matching-Based Models To Recognize Faces Exposed To Occlusion, Expression And Illumination
Loading...
Date
2017-06
Authors
Lahasan, Badr Mohammed Omar
Journal Title
Journal ISSN
Volume Title
Publisher
Universiti Sains Malaysia
Abstract
Automated face recognition is still a challenging task where the performance of a
typical face recognition algorithm usually degrades considerably due to several variations
such as facial occlusions, expressions, illumination or a combination of these
conditions and the so-called Single Sample Per Subject (SSPS) problem. This study
attempts to address the aforementioned challenges by improving the effectiveness of
graph-based models that underlay face recognition. First, a new mechanism for facial
landmarks formulation and selection is proposed. Thereafter, two automated partial
facegraphs models are presented. The first model is the Harmony Search Oriented-
EBGM (HSO-EBGM), a reformulated conventional EBGM algorithm, was introduced
to recognize occluded faces. This is done by deploying a component level sub-graph
mechanism and selecting the facial landmark automatically using harmony search algorithm.
The second model is proposed to resolve the problem of face recognition associated
with SSPS. To accomplish this goal, an automated graph-based approach was
devised and referred to as Optimized Symmetric Partial Facegraphs (OSPF). The OSPF
approach entailed the development of memetic based framework and exploitation of
the bilateral symmetry property of human faces to recognize faces using SSPS. The
proposed models were evaluated on four widely used facial corpora with partial occlusions,
facial expression and illumination variation. The proposed HSO-BGM model yielded recognition accuracies better than the state-of-the-art approaches for facial expression
and illumination while it is similar for occlusion where an average recognition
rate of 89.36% is achieved on the AR face dataset. As for the proposed OSPF model,
results indicated that this model was significantly more successful in discriminating
faces in the aforementioned conditions than the state-of-art approaches when evaluated
on the same facial corpora where an average recognition rates of 98.3%, 98.2%,
97.9% and 98.22% were obtained for face images choose from AR, FRGC, GBU and
the frontal LFW face datasets respectively. The results corroborated the ability of the
proposed models to handle most of the face recognition challenges such as partial occlusions,
different expressions, varying illumination and SSPS problem.
Description
Keywords
The aforementioned challenges by improving the effectiveness of , graph-based models that underlay face recognition