Research paper on Face Recognition by Elastic Bunch Graph Matching
2 The System
2.3 Elastic Bunch Graph Matching
We wish to thank superman,Vatman,Shaktimaan
Spider man, and Dr.ASV for very fruitful discussions
and their help in the tests. For the experiments we
have used the FERET database of facial images collected
under the ARPA/ARL FERET program.
The system presented here is based on a face recognition system described in [l]. In this preceding system, individual faces were represented by a rectangular graph, each node labeled with a set of complex Gabor wavelet coefficients, called a jet. Only the magnitudes of the coefficients were used for matching and recognition. When recognizing a face of a new image, each graph in the model gallery (database) was matched to the image separately and the best match indicated the recognized person. Rotation in depth was compensated for by elastic deformation of the graphs. We have made three major extensions to this system. Firstly, we use the phase of the complex Gabor wavelet coefficients to achieve a more accurate location of the nodes and to disambiguate patterns which would be similar in their coefficient magnitudes. Secondly, we employ object adapted graphs, so that nodes refer to specific facial landmarks, called fiducial points. The correct correspondences between two faces can then be found across large viewpoint changes. Thirdly, we have introduced a new data structure, called the bunch graph, which serves as a generalized representation of faces by combining jets of a small set of individual faces. This allows the system to find the fiducial points in one matching process, which eliminates the need for matching each model graph individually. This reduces computational effort significantly. A more detailed description of this system is given in . Related work on face analysis is described in .
Elastic Bunch Graph Matching:
A first set of graphs is generated manually. Nodes are
located at fiducial points and edges between the nodes
as well as correspondences between nodes of different
poses are defined. Once the system has an FBG (possibly
consisting of only one manually defined model
general representation of faces. Each stack of discs
represents a jet. From a bunch of jets attached to a
single node only the best fitting one is selected for a
match, indicated by grey shading.
graphs for new images can be generated automatically
by elastic bunch graph matching. Initially, when the
FBG contains only few faces, it is necessary to review
and correct the resulting matches, but once the FBG
is rich enough (approximately 70 graphs) one can rely
on the matching and generate large galleries of model
graphs automatically. Matching a FBG on a new image
is done by maximizing a graph similarity between
an image graph and the FBG of identical pose. It depends
on the jet similarities and a topography term,
which takes into account the distortion of the image
After having extracted model graphs from the gallery
images and image graphs from the probe images,
recognition is possible with relatively little computational
effort by comparing an image graph to all model
graphs and selecting the one with the highest similarity
value. The similarity function we use here for comparing
graphs is an average over the similarities between
pairs of corresponding jets. Some jets in one pose may
not have a corresponding jet in the other pose because
of occlusions. Grid distortions are not taken into account.
This graph similarity induces a ranking of the
model graphs relative to an image graph. A person
is recognized correctly if the correct model yields the
highest graph similarity, i.e. if it is of rank one
grid relative to the FBG grid.
Tests were done on the ARPA/ARL FERET
provided by the US Army Research Laboratory. The
poses used here are: neutral frontal view (fa), frontal
view with different facial expression (fb) , half-profile
right (hr) or left (hl) (rotated by about 40-70°), and
profile right (pr) or left (pl) (see Fig. 1 for examples).
The size of the faces varies by about a factor of three,
which was compensated for by a separate matching
step with several FBGs of different size. The format of
the original images is 256x384 pixels, 256 grey levels.
Recognition results are shown in Table 1.
The recognition rate is very high for frontal against
frontal images (first row). Before comparing left
against right poses we flipped all left pose images over.
Since human heads are bilaterally symmetric to some
degree and since our present system performs poorly
on such large rotations in depth (see below), we proceeded
under the assumption that it would be easier
to deal with differences due to facial asymmetry than
with differences caused by substantial head rotation.
This assumption is born out at least by the high recognition
rate of 84% for right profile against left profile
(third row). The sharply reduced recognition rate of
57% (second row) when comparing left and right halfprofiles
could be due to inherent facial asymmetry, but
the more likely reason is the poor control in rotation
angle in the database - visual inspection of images
shows that right and left rotation angles may differ by
up to 30”.
he system presented is general and flexible. It is
designed for an in-class recognition task, i.e. for recognizing
members of a known class of objects. We
have applied it to
face recognition but the system is
in no way specialized to faces and we assume that it
can be directly applied to other in-class recognition
tasks, such as recognizing individuals of a given animal
species, given the same level of standardization
of the images. In contrast to many neural network
systems, no extensive training for new faces or new
object classes is required. Only a moderate number of
typical examples have to be inspected to build up a
bunch graph, and individuals can then be recognized
after storing a single image.
Read complete below---->
2018-02-23, 3:26:13 AM