CUDA Accelerated Face Recognition

CUDA Accelerated Face Recognition
Numaan. A
Third Year Undergraduate
Department of Electrical Engineering
Indian Institute of Technology, Madras, India
ee08b044@smail.iitm.ac.in
Sibi. A
NeST-NVIDIA Centre for GPU Computing
NeST, India
sibi.a@nestgroup.net
July 26, 2010
Abstract
This paper presents a study of the efficiency and
performance speedup achieved by applying Graph-
ics Processing Units for Face Recognition Solutions.
We explore one of the possibilities of parallelizing and
optimizing a well-known Face Recognition algorithm,
Principal Component Analysis (PCA) with Eigen-
faces.
1 Introduction
In recent years, the Graphics Processing Units (GPU)
has been the subject of extensive research and the
computation speed of GPUs has been rapidly increas-
ing. The computational power of the latest genera-
tion of GPUs, measured in Flops1
, is several times
that of a high end CPU and for this reason, they
are being increasingly used for non-graphics applica-
tions or general-purpose computing (GPGPU). Tra-
ditionally, this power of the GPUs could only be
harnessed through graphics APIs and was primarily
used only by professionals familiar with these APIs.
1floating-point operations per second
CUDA (Compute Unified Device Architecture) de-
veloped by NVIDIA, tries to address this issue by in-
troducing a familiar C like development environment
to GPGPU programming and allows programmers to
launch hundreds of concurrent threads to run on the
“massively” parallel NVIDIA GPUs, with very little
software overhead. This paper portrays our efforts to
use this power to tame a computationally intensive,
yet highly parallelizable PCA based algorithm used in
face recognition solutions. We developed both CPU
serial code and GPU parallel code to compare the ex-
ecution time in each case and measure the speed up
achieved by the GPU over the CPU.
2 PCA Theory
2.1 Introduction
Principal Component Analysis (PCA) is one of the
early and most successful techniques that have been
used for face recognition. PCA aims to reduce the
dimensionality of data so that it can be economically
represented and processed. Information contained in
a human face is highly redundant, with each pixel
highly correlated to its neighboring pixels and the
1

main idea of using PCA for face recognition is to
remove this redundancy, and extract the features re-
quired for comparison of faces. To increase accuracy
of the algorithm, we are using a slightly modified
method called Improved PCA, in which the images
used for training are grouped into different classes
and each class contains multiple images of a single
person with different facial expressions. The mathe-
matics behind PCA is described in the following sub-
sections.
2.2 Mathematics of PCA
Firstly, the 2-D facial images are resized using bilin-
ear interpolation to reduce the dimension of data and
increase the speed of computation. The resized im-
age is represented as a 1-D vector by concatenating
each row into one long vector . Let’s suppose we have
M training samples per class, each of size N (=total
pixels in the resized image). Let the training vectors
be represented as xi. pj’s represent pixel values.
xi = [p1 . . . pN ]T
, i = 1, . . . , M
The training vectors are then normalized by sub-
tracting the class mean image m from them.
m =
1
M
M
i=1
xi
Let wi be the normalized images.
wi = xi − m
The matrix W is composed by placing the normal-
ized image vectors wi side by side. The eigenvectors
and eigenvalues of the covariance matrix C is com-
puted.
C = WWT
The size of C is N × N which could be enormous.
For example, images of size 16 × 16 give a covari-
ance of size 256 × 256. It is not practical to solve
for eigenvectors of C directly. Hence the eigenvectors
of the surrogate matrix WT
W of size M × M are
computed and the first M −1 eigenvectors and eigen-
values of C are given by Wdi and µi, where di and
µi are eigenvectors and eigenvalues of the surrogate
matrix, respectively.
The eigenvectors corresponding to non-zero eigen-
values of the covariance matrix make an orthonormal
basis for the subspace within which most of the image
data can be represented with minimum error. The
eigenvector associated with the highest eigenvalue re-
flects the greatest variance in the image. These eigen-
vectors are known as eigenimages or eigenfaces and
when normalized look like faces. The eigenvectors of
all the classes are computed similarly and all these
eigenvectors are placed side by side to make up the
eigenspace S.
The mean image of the entire training set, m
is computed and each training vector xi is normal-
ized. The normalized training vectors wi are pro-
jected onto the eigenspace S, and the projected fea-
ture vectors yi of the training samples are obtained.
wi = xi − m
yi = ST
wi
The simplest method for determining which class
the test face falls under is to find the class k, that
minimizes the Euclidean distance. The test image
is projected onto the eigenspace and the Euclidean
distance between the projected test image and each
of the projected training samples are computed. If
the minimum Euclidean distance falls under a prede-
fined threshold θ, the face is classified as belonging to
the class to which contained the feature vector that
yielded the minimum Euclidean distance.
3 CPU Implementation
3.1 Database
For our implementation, we are using one of the most
common databases used for testing face recognition
solutions, called the ORL Face Database (formerly
known as the AT&T Database). The ORL Database
2

contains 400 grayscale images of resolution 112 ×
92 of 40 subjects with 10 images per subject. The
images are taken under various situations, such as
different time, different angles, different expressions
(happy, angry, surprise, etc.) and different face de-
tails (with/without spectacles, with/without beard,
different hair styles etc). To truly show the power
of a GPU, the image database has to be large. So
in our implementation we scaled the ORL Database
multiple times by copying and concatenating, to cre-
ate bigger databases. This allowed us to measure the
GPU performance for very large databases, as high
as 15000 images.
3.2 Training phase
The most time consuming operation in the training
phase is the extraction of feature vector from the
training samples by projecting each one of them to
the eigenspace. The computation of eigenfaces and
eigenspace is relatively less intensive since we have
used the surrogate matrix workaround to decrease the
size of the matrix and compute the eigenvectors with
ease. Hence we have decided to parallelize only pro-
jection step of the training process. The steps till the
projection of training samples are done in MATLAB
and the projection is written in C.
The MATLAB routine acquires the images from
files, resizes them to a standard 16 × 16 resolution
and computes the eigenfaces and eigenspace. The
data required for the projection step, namely, the
resized training samples, the database mean image
and the eigenvectors, are then dumped into a bi-
nary file with is later read by the C routine to com-
plete the training process. The C routine reads the
data dumped by the MATLAB routine and extracts
the feature vectors by normalizing the resized train-
ing samples and projecting each of them onto the
eigenspace. With this the training process is com-
plete and the feature vectors are dumped onto a bi-
nary file to be used in the testing process.
3.3 Testing phase
The entire testing process is written in C. OpenCV
is used to acquire the testing images from file, and
resize them to the standard 16 × 16 resolution using
bilinear interpolation. The resized image is normal-
ized with the database mean image and projected
onto the eigenspace computed in the training phase.
The euclidean distance between the test image fea-
ture vector and the training sample feature vectors
are computed and the index of the feature vector
yielding the minimum euclidean distance is found.
The face that yielded this feature vector is the most
probable match for the input test face.
4 GPU Implementation
4.1 Training phase
4.1.1 Introduction
As mentioned in Section 3.2, only the final feature ex-
traction process of the training phase is parallelized.
Before the training samples can be projected, all the
data required for the projection process is copied to
the device’s global memory and the time taken for
copying is noted. As all the data are of a read-only
nature, they are bound as texture to take advantage
of the cached texture memory.
4.1.2 Kernel
The projection process is highly parallelizable and
can be parallelized in two ways. The threads can be
launched to parallelize the computation of a partic-
ular feature vector, wherein, each thread computes a
single element of the feature vector. Or, the threads
can be launched to parallelize projection of multiple
training samples, wherein each thread projects and
computes the feature vector of a particular training
sample. Since the number of training samples is large,
the latter is adopted for the projection operation in
training phase. We have adopted the former in the
testing phase, where only one image has to be pro-
jected, details of which are explained in Section 4.2.
Before the projection kernel is called, the execution
configuration is set. The number of threads per block,
T1, is set to a standard of 256 and the total number of
blocks is B1, where B1 = (ceil) (N1/T1), N1 = total
number of training samples.
3

Each thread projects and computes the feature vec-
tor of a particular training sample by serially comput-
ing each element of the feature vector one by one.
Each element of the feature vector is obtained by
taking inner product of the training image vector
and the corresponding eigenvector in the eigenspace.
The training sample is normalized with the database
mean image element by element as it is fetched from
texture memory and the intermediate sum of the in-
ner product with eigenvector is stored in the shared
memory. After each element of the feature vector is
computed, the data is written back into the global
memory and the next element is computed. All the
data is aligned in a columnar fashion to avoid un-
coalesced memory accesses and shared memory bank
conflicts.
Figure 1: Threads in Projection Kernel (Training)
After the kernel has finished running on the device
the entire data is copied back to the host memory
and dumped as a binary file to be used in the testing
phase.
4.2 Testing Phase
4.2.1 Introduction
The testing process is completely run on the GPU
and is handled by three kernels. The first kernel
normalizes and projects it onto the eigenspace and
extracts the feature vector. The second kernel par-
allely computes the euclidean distance between the
feature vector of the test image and that of the train-
ing images. The final kernel, finds the minimum of
the euclidean distance and index of the training sam-
ple which yielded that minimum. The resized test
image, database mean image, eigenvectors and the
projected training samples are first copied to the de-
vice memory and the test image, mean image and
eigenvectors are bound as texture to take advantage
of the cached texture memory. Due to the relatively
larger size of the projected training samples data and
the limitation on maximum texture memory, the pro-
jected training samples are not bound as texture.
Figure 2: Recognition pipeline
4.2.2 Projection Kernel
As mentioned in section 4.1.2, the projection kernel
in testing process is parallelized to concurrently com-
pute each element of the feature vector. The number
of threads per block, T2, is set to a standard of 256
and the total number of blocks is B2, where B2 =
(ceil) (N2/T2), N2 = size of feature vector.
Each thread computes each element of the feature
vector, which is obtained by taking inner product of
the test image vector and the corresponding eigenvec-
tor in the eigenspace. The test image is normalized
with the database mean image, element by element
as it is fetched from texture memory and the inter-
mediate sum of the inner product with eigenvector is
4

stored in the shared memory.
Figure 3: Threads in Projection Kernel (Testing)
After the entire feature vector is computed, the
data is written back into global memory. The colum-
nar alignment of all the eigenvectors avoids unco-
alesced memory accesses and shared memory bank
conflicts.
4.2.3 Euclidean Distance Kernel
The kernel for computing the euclidean distance is
very similar to the projection kernel used in training
phase. Threads are launched to concurrently com-
pute the euclidean distance between the test image
feature vector and the training sample feature vec-
tors. The number of threads per block, T3, is set to
a standard of 256 and the total number of blocks is
B3, where B3 = (ceil) (N3/T3), N3 = total number
of training samples.
Each thread computes a particular euclidean dis-
tance serially. The difference of each element of the
vectors are computed, squared and summed. The in-
termediate sum is stored in shared memory. After all
the euclidean distances are computed, the data from
the shared memory is written to the global mem-
ory. The columnar alignment of training sample fea-
ture vectors avoids uncoalesced memory accesses and
shared memory bank conflicts.
Figure 4: Threads in Euclidean Distance Kernel
4.2.4 Minimum Kernel
The minimum kernel computes the minimum value
of euclidean distance and its distance. The vector
containing euclidean distances is divided into smaller
blocks and each thread serially finds the value and
index of the minimum in a particular block. The ker-
nel is called iteratively with fewer and fewer threads
till only one block is left. After execution of the ker-
nel the minimum value and its index is copid back
to host memory. The training sample at the index
computed by the kernel is the most probable match
for the test image.
5 Performance Statistics
To test the performance of CPU and GPU, 5 images
per subject of the ORL Database was selected as the
training set. This set of 200 images was then repli-
cated and concatenated to create databases of size
ranging from 1000 to 15000 images. The eigenvectors
corresponding to the 4 highest eigenvalues per class,
were selected for forming the eigenspace. This led
to feature vectors which grew in size as the database
grew. This replicated database was trained with CPU
and GPU and the execution time was noted and the
GPU speedup for the training process was calculated.
5

For testing, one image per subject from the ORL
Database were selected and the total time taken by
the CPU and GPU to test all 40 test images was
noted and was used to calculate the GPU speedup
for testing process.
To get accurate performance statistics, the training
and testing processes were run multiple times on dif-
ferent CPUs and GPUs. The following graphs were
plotted with the data obtained from the performance
tests. All the CPU times are based on single-core
performances.
Fig. 5 shows the time taken three different CPUs
to execute the projection process during training.
Figure 5: Training time for different CPUs
Fig. 6 shows the time taken by different NVIDIA
GPUs to execute the projection process during train-
ing. It includes the time taken for data transfers to
and from the device.
Figure 6: Training time for different GPUs
Fig. 7 shows the performance speedup of different
GPUs over Intel Core 2 Quad Q9550 CPU during
training databases of varying sizes.
Figure 7: Training Speedup
Fig. 8 shows total time taken by different CPUs
for testing 40 images.
6

Figure 8: Testing using GPU
Fig. 9 shows total time taken by GPUs to test 40
images. For this test, the trained database was copied
to device memory once and 40 images were tested
one by one. It includes time taken for transferring
test image to device and getting match index from
device.
Figure 9: Testing using GPU
Fig. 10 shows the performance speedup of diﬀer-
ent GPUs over Intel Core 2 Quad Q9550 CPU when
testing 40 images.
Figure 10: Testing Speedup
Fig 11 shows the execution time of the recognition
pipeline on the GPU for varying database sizes.
Figure 11: Recognition pipeline on CPU
7

Fig. 12 shows the execution time of the recognition
pipeline on the GPU for varying database sizes. It is
the time taken to transfer test image to device, ﬁnd
the match and transfer its index back to host.
Figure 12: Recognition Pipeline on GPU
Fig. 13 shows the performance speedup of diﬀer-
ent GPUs over Intel Core 2 Quad Q9550 CPU when
executing the recognition pipeline.
Figure 13: Recognition Pipeline Speedup
6 Conclusion
The recognition rate of a PCA based face recognition
solution depends heavily on the exhaustiveness of the
training samples. Higher the number of training sam-
ples, higher the recognition rate. But as the num-
ber of training samples increases, CPUs get highly
strained and the training process will take several
minutes to complete (refer Fig. 5). But the same
process, when run on a GPU, will be completed in
a manner of seconds (refer Fig. 6). The highest
speedup achieved was 207x for training process, 330x
for the recognition pipeline and 165x for overall test-
ing process on the latest GeForce GTX 480 GPU, for
a database size of 15,000 images.
The execution time of the recognition pipeline on
the GPU is in the order of a few milli seconds even
for very large databases and and this allows the GPU
based testing to be integrated with real time video
and used for other applications involving large vol-
umes of test images. Our primary purpose in writ-
ing this paper is to make clear, the high perfor-
mance boosts that can be obtained by developing
GPU based face recognition solutions.
8

7 Future Works
Our future plans on this ﬁeld include the paralleliza-
tion of other face recognition algorithms like LDA
(Linear Discriminant Analysis) and to replace the eu-
clidean distance based matching process with a neu-
ral network based one. We feel that algorithms with
a high degree of parallelism in them, like neural net-
works, will beneﬁt immensely, if implemented on the
GPU. We are also working on integrating the GPU
recognition pipeline with real time video.
9

CUDA Accelerated Face Recognition

More Related Content

What's hot (18)

Viewers also liked (16)

Similar to CUDA Accelerated Face Recognition (20)

More from QuEST Global (erstwhile NeST Software) (8)

Recently uploaded (20)

CUDA Accelerated Face Recognition