Facial recognition using composite correlation filters designed
with multiobjective combinatorial optimization
Andres Cuevas
a
, Victor H. Diaz-Ramirez
a
, Vitaly Kober
b,c
, and Leonardo Trujillo
d
a
Instituto Polit´ ecnico Nacional - CITEDI, Ave. del Parque 1310, Mesa de Otay, Tijuana B.C.
22510, Mexico;
b
Department of Computer Science, CICESE, Carretera Ensenada-Tijuana 3918, Ensenada
B.C. 22860, Mexico;
c
Department of Mathematics, Chelyabinsk State University, Russian Federation;
d
Instituto Tecnologico de Tijuana, Blvd. Industrial y Ave. ITR, Tijuana S/N, Mesa de Otay,
Tijuana B.C. 22414, Mexico
ABSTRACT
Facial recognition is a difficult task due to variations in pose and facial expressions, as well as presence of noise and
clutter in captured face images. In this work, we address facial recognition by means of composite correlation
filters designed with multi-objective combinatorial optimization. Given a large set of available face images
having variations in pose, gesticulations, and global illumination, a proposed algorithm synthesizes composite
correlation filters by optimization of several performance criteria. The resultant filters are able to reliably detect
and correctly classify face images of different subjects even when they are corrupted with additive noise and
nonhomogeneous illumination. Computer simulation results obtained with the proposed approach are presented
and discussed in terms of efficiency in face detection and reliability of facial classification. These results are also
compared with those obtained with existing composite filters.
Keywords: Facial recognition, composite correlation filters, multi-objective evolutionary algorithm, combina-
torial optimization.
1. INTRODUCTION
Over the last decades correlation filters have been successfully used for object detection and pattern classification
problems.
1
A correlation filter is a linear system where the coordinates of the system’s output maximum are
position estimates of the target within the scene. One advantage of correlation filters is that they perform well
under noisy conditions.
2–4
Also, they possess a degradation-invariant property that tolerates partial occlusions of
the target without significantly sacrificing their performance.
5
Correlation filters can be broadly classified in two
main categories: analytical and composite. Analytical filters optimize a performance criterion from mathematical
models of signals and noise.
6, 7
For facial recognition problems, these filters may not be recommended because the
inherent variations in facial expression will rapidly degrade their recognition performance. On the other hand,
composite filters are constructed by combination of several training templates, each of them representing an
expected view of the target in the input image.
5
These filters have been successfully used for facial recognition.
1
Composite filters are synthesized by means of a supervised training process. Thus, the performance of the filters
highly depend on a proper selection of templates used for training.
8
Commonly, training templates are chosen
by a designer in an improvised manner. So, this subjective procedure is not optimal.
In order to synthesize composite filters with improved performance for distortion tolerance facial recognition,
we propose a filter design algorithm based on multi-objective combinatorial optimization. Given a large search
space of training templates, the algorithm finds a subset of these templates that allows the synthesis of com-
posite filters with an optimized performance in terms of several competitive criteria. The proposed algorithm
combines two evolutionary strategies. Firstly, population management is performed by the Strength Pareto Evo-
lutionary Algorithm 2 (SPEA2),
9
which is a state-of-the-art multi-objective evolutionary algorithm (MOEA).
Secondly, individuals of the population are represented as variable length strings, while genetic operators applied
to individuals are managed by the Speciation Adaptation Genetic Algorithm (SAGA).
10
The proposed design
Applications of Digital Image Processing XXXVII, edited by Andrew G. Tescher, Proc. of SPIE Vol. 9217, 921710
© 2014 SPIE · CCC code: 0277-786X/14/$18 · doi: 10.1117/12.2062348
Proc. of SPIE Vol. 9217 921710-1
Downloaded From: http://proceedings.spiedigitallibrary.org/ on 09/24/2014 Terms of Use: http://spiedl.org/terms