225
Abstract
Recently sparse and collaborative representation based
classification has been developed for face recognition with
single sample per person (SSPP). By using variations
extracted from a generic training set as an additional
common dictionary, promising performance has been
reported in face recognition with SSPP. However, existing
representation based classifiers for face recognition with
SSPP ignored to make full use of the discrimination of
particular facial regions (e.g., eyes and nose), which have a
high detection rate in various facial variations. Since the
particular regions only cover a certain part of the face, the
regular regions (e.g., dense sampling points) are also
utilized to represent the face completely. In this paper we
proposed a robust local representation (RLR) model for
face recognition with SSPP by fully exploiting the features
of both particular and regular facial regions. Thus
variation-robust features of particular facial regions and
dense features of regular facial regions are both captured,
while bad features with big variations are adaptively given
low weight values by the proposed RLR so that they
contribute little to the representation and classification.
Experimental results on AR and LFW databases
demonstrate that the proposed RLR method has achieved
better results than the state of the art methods.
1. Introduction
As one of the most visible applications in computer
vision and pattern recognition, face recognition (FR) has
been receiving significant attention in the community [17].
In practical FR scenarios such as face
identification/verification in uncontrolled or less controlled
environment [6, 16], there are many problems which have
attracted much attention of researchers. For instance, face
recognition with single sample per person is one of the most
important FR problems. In the scenarios (e.g., law
enforcement, e-passport, driver license, etc.), there is only a
single training face image per person. This makes the
problem of FR particularly hard since very limited
information is provided to predict the variations in the query
sample. How to achieve high FR performance in the case of
single training sample per person (SSPP) is an important
and challenging problems in FR.
The performance of FR would be greatly affected by the
limited number of training samples per person [26]. First,
many discriminative subspace and manifold learning
algorithms (e.g., LDA and its variants [15]) cannot be
directly applied to FR with SSPP. Second, sparse
representation based classification (SRC) [12], cannot be
easily applied to the problem of SSPP, either, since multiple
training samples per person are needed to well reconstruct
the query face. As reviewed in [26], many specially
designed FR methods have been developed. According to
the availability of an additional generic training set, the FR
methods for SSPP can be divided into two categories:
methods without using a generic training set, and methods
with generic learning.
The SSPP methods without generic learning often extract
robust local features (e.g., gradient orientation [10] and
local binary pattern [1]), generate additional virtual training
samples (e.g., via singular value decomposition [25],
geometric transform and photometric changes [27]), or
perform image partitioning (e.g., local patch based LDA
[23], self-organizing maps of local patches [22], and
multi-manifold learning from local patches [8]). Although
these methods have reported improved FR results, they
ignored to introduce additional variation information into
the single-sample gallery set. Meanwhile the new
information introduced by virtual training sample
generation can be rather limited.
Opposite to the first category of FR with SSPP, methods
with generic learning try to borrow new and useful
information (e.g., generic intra-class variation) from a
generic training set. An intrinsic reason is the fact that face
image variations for different subjects share much similarity.
Since a generic training set could be easily collected, it has
been widely employed in [9, 20, 21] to extract
discriminative information for FR with SSPP. For instance,
the expression subspace and pose-invariant subspace were
learned from a collected generic training set to solve the
expression-invariant [21] and pose-invariant [9] FR
problems, respectively. Deng et al. [3] extended the SRC
Robust Local Representation for Face Recognition with Single Sample Per Person
Xing Wang
1
Meng Yang
1
* Linlin Shen
1
Heyou Chang
2
1
College of Computer Science and Software Engineering, Shenzhen University
2
School of Computer Science and Technology, Nanjing University of Science and Technology
*Corresponding Author; yang.meng@szu.edu.cn