30-11-2012, 04:04 PM
Enhanced Patterns of Oriented Edge Magnitudes for Face Recognition and Image Matching
1Enhanced Patterns of Oriented.pdf (Size: 1.41 MB / Downloads: 60)
Abstract
A good feature descriptor is desired to be discriminative,
robust, and computationally inexpensive in both terms of time
and storage requirement. In the domain of face recognition, these
properties allow the system to quickly deliver high recognition
results to the end user. Motivated by the recent feature descriptor
called Patterns of Oriented Edge Magnitudes (POEM), which
balances the three concerns, this paper aims at enhancing its
performance with respect to all these criteria. To this end, we first
optimize the parameters of POEM and then apply the whitened
principal-component-analysis dimensionality reduction technique
to get a more compact, robust, and discriminative descriptor.
For face recognition, the efficiency of our algorithm is proved by
strong results obtained on both constrained (Face Recognition
Technology, FERET) and unconstrained (Labeled Faces in the
Wild, LFW) data sets in addition with the low complexity. Impressively,
our algorithm is about 30 times faster than those based on
Gabor filters. Furthermore, by proposing an additional technique
that makes our descriptor robust to rotation, we validate its
efficiency for the task of image matching.
INTRODUCTION
F ACE recognition is an established field with a wide variety
of real-world applications. Generally, within the domain
of face recognition, facial feature description is the most important
aspect. If inadequate features are used, even the best classifier
will fail to achieve good recognition results. A good face descriptor
is desired to have the following properties: 1) minimize
intraperson dissimilarities; 2) enlarge the margin between different
people (this is a critical issue, as variations of pose, illumination,
age, and expression can be larger than variations of identity);
and 3) can be easily extracted from original data to allow
fast processing, as well as lie in a low-dimensional space to
allow both inexpensive storage requirements and fast classification
stage. These properties allow the system to quickly deliver
highly accurate results to the end user.
RELATED WORK
The primary motivation of this work is finding an efficient
facial representation for face recognition, and thus, this section
mainly discusses the related literature on face descriptors.
Of course, regarding the application of our algorithm to image
matching, in order to make this paper more self-contained, a
brief overview of such application is also provided.
Face Representation
There are many representational approaches used within the
domain of face recognition, including subspace-based holistic
features, such as eigenface [4] and fisherface [5], and local appearance
features [6], [7]. Heisele et al. [8] compared local and
global approaches, and observed that local systems outperform
global systems for recognition rates larger than 60%. Consequently,
this section highlights only local face descriptors. However,
before going into the details, note that, in our consideration,
we make a distinction between elementary feature descriptors,
which return the feature vector using only the image itself,
and learned approaches, which require a training set for learning
the descriptors. Learning techniques range from classical dimensionality
reduction algorithms such as PCA and linear discriminant
analysis (LDA), or feature selection methods such as
Adaboost to the more recently proposed methods, such as Bag
of Feature [9].
Two of the most successful local face representations are
Gabor features and Local Binary Patterns (LBPs). Gabor
features, which are spatially localized and selective to spatial
orientations and scales, are analogous to the receptive fields of
simple cells in the mammalian visual cortex [7]. Due to their
robustness to local distortions, Gabor features (or variants, e.g.,
Gabor jets) have been successfully applied to face recognition
[7], [10]–[15]. Indeed, the FERET evaluation and FRGC2004
contests have seen top performance from methods based on
Gabor features.
POEM
The current state-of-the-art face representations are those
combining or fusing different single features: LBP and Gabor
features as in [10] and [11], and LBP, Gabor, and SIFT as in
[24] and [35]. These algorithms try to bring the advantages of
different single features: the LBP method is a “micropattern”
capturing image details at fine scales, whereas the Gabor filters
are capable of characterizing image information over coarser
scales and through different orientations. While the LBP
method is a good choice for describing texture information,
the SIFT and HOG approaches are widely accepted as the best
features to capture edge or local shape information.
In order to build features that can inherit these properties
without suffering the shortcoming of the Gabor filters, i.e., the
computationally expensive cost, we propose to apply the idea
of self-similarity calculation from the LBP-based structure on
the distribution of local edge through different orientations. The
resulting features are referred to POEM. More precisely, to calculate
the POEM codes for one pixel, the intensity values in
the calculation of conventional LBP are replaced by gradient
magnitudes, which are calculated by accumulating a local histogram
of gradient directions over all pixels within a spatial
patch (“cell”). Additionally, these calculations are done across
different orientations.
CONCLUSION
By applying the self-similarity operator on accumulated edge
magnitudes across different directions, we have developed a
novel feature set called POEM for feature extraction. Based on
these features, two feature descriptors are built, i.e., POEM-HS
and WPCA-POEM. Both algorithms have several desirable features:
robust to lighting, pose, and expression variations, and
is fast to compute when compared to many of the competing
descriptors. We have shown that they are very effective representations
for face recognition in both constrained (FERET)
and unconstrained (LFW) face recognition tasks, not only outperforming
all other descriptor-based techniques but also being
comparable with other much more complex systems. The high
performance coupled with very low complexity in both terms of
computational time and storage requirements suggests that our
algorithm is a good candidate for use in real-world face recognition
systems, which must quickly deliver high-quality results to
the end user. Furthermore, by presenting an additional technique
that makes our descriptor robust to rotation, we demonstrate its
efficiency for the task of image matching.