Face detection on_embedded_systems

Face Detection on Embedded Systems
Abbas Bigdeli1, Colin Sim2, Morteza Biglari-Abhari2 and Brian C. Lovell1
1

Safeguarding Australia Program,
NICTA, Brisbane, QLD 4000, Australia
{abbas.bigdeli, brian.lovell}@nicta.com.au
2
Department of Electrical and Computer Engineering
The University of Auckland, Auckland, New Zealand
{csim036, m.abhari}@auckland.ac.nz

Abstract. Over recent years automated face detection and recognition (FDR)
have gained significant attention from the commercial and research sectors.
This paper presents an embedded face detection solution aimed at addressing
the real-time image processing requirements within a wide range of
applications. As face detection is a computationally intensive task, an
embedded solution would give rise to opportunities for discrete economical
devices that could be applied and integrated into a vast majority of applications.
This work focuses on the use of FPGAs as the embedded prototyping
technology where the thread of execution is carried out on an embedded softcore processor. Custom instructions have been utilized as a means of applying
software/hardware partitioning through which the computational bottlenecks
are moved to hardware. A speedup by a factor of 110 was achieved from
employing custom instructions and software optimizations.

1 Introduction
The identification and localization of a face or faces from either an image or video stream
is a branch of computer vision known as face detection [1, 2]. Face detection has attracted
considerable attention over recent years in part due to the wide range of applications in
which it forms the preliminary stage. Some of the main application areas include: humancomputer interaction, biometrics, content-based image retrieval systems (CBIRS), video
conferencing, surveillance systems, and more recently, photography.
The existing visual sensing and computing technologies are at a state where reliable,
inexpensive, and accurate solutions for non-intrusive and natural means of humancomputer interactions are feasible. Biometrics is an evolving application domain for face
detection and is concerned with the use of physiological information to identify and verify
a person’s identity. In most cases, face recognition algorithms are designed to operate on
images assumed to only contain frontal faces [2]. Therefore, face detection is required to
first extract faces from an image prior to the recognition step. Examples of commercial
biometric systems are BioID [3] and ViiSage1. HumanScan is the company that developed
BioID; a multimodal system incorporating voice, lip movement and face recognition to

1

www.viisage.com

2

Abbas Bigdeli, Colin Sim, Morteza Biglari-Abhari and Brian C. Lovell

authenticate a person. This system implements a model-based face detection algorithm
based on the Hausdorff distance [4].
Another application area that can clearly benefit from face detection is surveillance
systems that would allow easier identification of criminals in public spaces. Shan et al [5]
presented a robust face recognition system specifically designed for Intelligent CCTV
systems. Another video surveillance system which has the capacity to detect faces is that
proposed by Kim et al [6]. More recently, FujiFilm2 and Nikon Corporation3 have
incorporated face detection technologies into some of their camera series to automatically
improve pictures taken under poor lighting conditions.
The majority of the research work to date has primarily focused on developing novel
face detection algorithms and/or improving the efficiency and accuracy of existing
algorithms. As a result, most solutions deployed (similar to the examples given above) are
typically high-level software programs targeted for general purpose processors that are
expensive and usually non-real-time solutions. Since face detection is typically the first
step and frequently a bottleneck in most solutions due to the large search space and
extensive amount of computationally intensive operations, it is reasonable to suggest an
embedded implementation specifically optimized to detect faces. An embedded solution
would entail many advantages including 1) low cost, as only a subset of hardware
components are required compared to the general computer based solutions, 2)
optimization of the face detection algorithms for real-time operations independent of face
recognition or other post-processing concerns and 3) integration with other technologies
such as security cameras to create smart devices.
Related Work
Now that reliable, accurate, and efficient face detection algorithms are available coupled
with advances in embedded technologies; low-cost implementations of robust real-time
face detectors can be explored. The most common target technologies are: pure hardware,
embedded microprocessors, and configurable hardware.
Pure hardware systems are typically based on very large scale integrated circuit (VLSI)
semiconductor technology implemented as application specific integrated circuits (ASIC).
Compared to the other technologies, ASICs have a high operating frequency resulting in
better performance, low power consumption, high degree of parallelism, and well
established design tools. However, a large amount of development time is required to
optimize and implement the designs. Also, due to the fixed nature of this technology the
resulting solutions are not flexible and cannot be easily changed, resulting in high
development costs and risk. Theocharides et al [7] investigated the implementation of a
neural network based face detection algorithm in 160 nm VLSI technology based on
algorithm proposed by Rowley et al in [8, 9], which has a high degree of parallelism.
On the other hand, software programs implemented on general purpose processors
(GPP) offer a great deal of flexibility, coupled with very well established design tools that
can automatically optimize the designs with little development time and costs. GPPs are
ideally suited to applications that are primarily made up of control processing. However,
they are disadvantaged because minimal or no special instructions are available to assist
with data processing [10]. Digital signal processors (DSP) extend GPPs in the direction of
increasing parallelism and providing additional support for applications requiring large
amounts of data processing. The drawbacks of microprocessors (both GPPs and DSPs) are
2
3

www.fujifilmusa.com
www.nikonimagings.com


3

high power consumption, and inferior performance compared to an ASIC. The
performance of the final solution is limited to the selected processor.
Finally, configurable platforms such as field programmable gate arrays (FPGA)
combine some of the advantages from both pure hardware and pure software solutions.
More specifically, the high parallelism and computational speed of hardware, and the
flexibility and short design time of software. By inheriting characteristics from both
hardware and software solutions, the design space for FPGAs is extended for better tradeoffs between performance and cost. These design trade-offs are far superior to that of pure
hardware or software solutions alone. From an efficiency point of view, the performance
measures for FPGAs, that is, operating frequency, power consumption, and so on, are
generally half way in between the corresponding hardware and software measures.
Several configurable hardware based implementations exist, including that by
McCready [11] and Sadri et al [12]. McCready specifically designed a novel face
detection algorithm for the Transmogrifier-2 (TM-2) configurable platform. The
Transmogrifer-2 is a multi-board FPGA based architecture proposed by Lewis et al [13].
The algorithm was intentionally designed with minimal mathematical operations that
could execute in parallel — engineering effort has been put in to reduce the number of
multiplications required. The implemented system required nine boards of the TM-2
system, requiring 31,500 logic cells (LC). The system can process 30 images per second
with a detection accuracy of 87%. The hardware implementation is said to be 1,000 times
faster than the equivalent software implementation.
On the other hand, Sadri et al [12] implemented the neural network based algorithm
proposed by Rowley et al [8] on the Xilinx Virtex-II Pro XC2VP20 FPGA. Skin color
filtering and edge detection is incorporated to reduce the search speed. The solution is
partitioned such that all regular operations are implemented in hardware while all irregular
control based operations are implemented on Xilinx’s embedded hardcore PowerPC
processor. This partitioning allows the advantages of both hardware and software to be
simultaneously exploited. The system operates at 200 MHz and can process up to nine
images per second.
The examples presented illustrate the obvious compromises between accuracy and
algorithm robustness versus the amount of resources required. That is, to improve the
performance of the face detection algorithms, we must either increase the embedded
design complexity, which generally results in higher power consumption and hardware
costs, or settle for a lesser solution.

2. PC Based Software Prototype
The initial software prototype of the standard Viola-Jones algorithm was implemented
based on the trained classifiers provided in the Open Computer Vision Library
(OpenCV)4. The particular classifiers used in this implementation are those trained for a
base detection window size of 24x24 pixels. The classifiers are trained to detect upright
frontal faces. These classifiers were created and trained by Lienhart et al who used a total
of 8,000 training samples, of which 5,000 are face images and 3,000 non-face images [14].
The accuracy of the implemented Viola-Jones face detection algorithm was validated
using a subset of images from the CMU + MIT face databases, and images retrieved from
a random web crawl.
4

www.opencv.org

4


To ensure that the results obtained can be compared with other published sources,
detection and false detection events use the definition of Lienhart et al [14], as follows. A
detection window is said to correctly identify a face if the following criteria are satisfied:
i) The maximum displacement between the centre of the detected window and the
actual face do not exceed 30% of the actual face size.
ii) The difference between the detected window and actual face size does not exceed
50% of the actual face size.
At ∆ = 1.0 and s = 1.255, the detection accuracy for the CMU+MIT image set is 80%. This
result is consistent with that presented by Lienhart et al [14]. On the other hand, the
detection accuracy for the web crawl image set is 92%. The slightly better results obtained
in the web crawl image set is possibly due to the following factors:
i) The images from the web crawl image set had a higher resolution and better quality
ii) The images’ backgrounds were less complex, resulting in less misclassification.
2.1 Implementation Details
In order to standardize the input data for all implementation platforms so that the
performance results can be benchmarked, a set of 10 images arbitrarily chosen from a web
search is used as input. Each image contains a single frontal face. Multiple images are
used so that any variability in the processing time for individual images is better averaged
out. All the images are stored as grayscale bitmaps of size 576x720 pixels — image size
being arbitrarily chosen. Variability in execution times is primarily attributed to the
cascaded nature of the Viola-Jones algorithm. The amount of time required to search an
image for a face is closely related to the “complexity” of the image; that is, if an image has
large areas that do not contain faces but passes many of the cascaded classifier stages, then
more time is required to process the image. Other aspects that contribute to varying the
execution times are platform dependent and include: cache and memory access times,
pipeline schedule, interrupt mechanisms, and so on.
The software prototype was developed in C and compiled with a GNU GCC compiler
under Cygwin. The compiler was set to highest optimization for speed, and the PC was a
Pentium 4 processor, 3.20 GHz, with 1.0 GB RAM with Windows XP as the operating
system.
2.2 Software Implementation Results
All 10 faces within the 10 images were located with two false positives. Due to the small
number of images, the ratio between detection and false detection rates may be skewed.
The total execution time for the program to process all 10 images is 19.91 seconds. A
breakdown of each function and their contribution to the final time are presented in the
Table 1.
Not all functions used in the program are in Table 1; this is because the profiling
process is based on samples taken when the program is running. As a result, functions that
execute quickly may not be acknowledged. An arrow before a function name indicates that
it is called from another function where the function that called it has one less indentation
level. As an example, referring to Table 1, InitialiseClassifierStages,
UpdateClassifierStages, and so on are all called from the FaceDetection
function. As seen in the flat profile, the RunClassifierCascade function consumes
the majority of the execution time. This function is responsible for checking if a sub5

Where s is the scale factor in the search algorithm and, ∆ represents the number of pixels to
shift the detection window [14]


5

window contains a face. The function itself is executed in close to no time; the long
processing time is due to the accumulation of over 7.5 million calls. The number of calls
to RunClassifierCascade is related to the number of search locations which in turn is
proportional to the size of the image.
Table 1. Flat profile of the Viola-Jones algorithm.
Function Name
CalculateIntegralImage
CalculateSqIntegralImage
FaceDetection
→ InitialiseClassifierStages
→ UpdateClassifierStages
→ RunClassifierCascade
→ round
→ AddSubWindow
→ GetFaceWindows
FreeIntegralImage
FreeFaceWinList
FreeSubWinList

Total Time Taken Number of Average Time Percentage of
(sec)
Calls
(sec)
Total Time (%)
0.35
10
0.04
1.76
0.03
10
0
0.15
19.53
10
1.95
98.09
0.02
10
0
0.1
0.02
150
0
0.1
19.46
7,458,250
0
97.74
0.01
Unknown
N/A
0.05
0
110
0
0
0
10
0
0
0
20
0
0
0
10
0
0
0
10
0
0

2.3 Initial Embedded System and Port of Software Code
A fully functional embedded system based on Altera Nios II softcore processor was
created on a Stratix FPGA development board. Overall the maximum system operating
frequency is 96.42 MHz. However, due to the limitations of the possible frequencies that
can be generated by the PLL module, the maximum feasible operating frequency without
violating timing constraints is 80 MHz. From this point onwards unless otherwise stated
the actual operating frequency for each system is 80 MHz.
Given that the code implemented on the PC is based on GNU libraries which is the
same as the software development in the Nios II environment, it was possible to be
directly ported across to the Nios II system without many changes.
2.4 Profile of Embedded Code
Initially, the GNU profiler was used to profile the code which was compiled with
maximum speed optimization. The original code runs quite slowly. The lengthy
processing time is attributed to the extensive number of locations searched within each
image. The amount of searching is directly proportional to the input image size and the
step size of the detection window. There are several methods available to help reduce the
search space, common approaches being, image differencing and skin color modeling.
Image differencing as its name suggests involves taking the difference between image
frames. Since humans are generally the only objects that move within a scene, searching
can be concentrated on the sections that are changing. On the other hand, the skin color
modeling method makes use of statistical models of human skin to filter the input image
hence the search for faces can then be restricted to those areas. An alternative method is to
reduce the size of the input images using an image interpolation algorithm. Once the faces
are found in the smaller image, the location and dimension of the detection windows can
be rescaled to the size of the original image.
The most popular techniques for enlarging or shrinking images are nearest
neighborhood, bilinear, and bicubic down-sampling. Of these methods, bicubic offers the
best results in terms of sharpness and preservation of image details. Every point in the

6


resulting image is calculated from a weighted average of 16 pixels adjacent to the
corresponding pixel in the original image. The bicubic interpolation algorithm can be
expressed as the following set of equations:
g ( x) = ∑ c k u ( s k )
k

⎧ A1 s 3 + B1 s 2 + C 1 S + D1
⎪
3
2
⎪
u ( s ) = ⎨ A2 s + B2 s + C 2 S + D2
⎪
0
⎪
⎩
Bicubic interpolation function and its kernel [15].

(1)

Where g is the interpolation function, c is points in the original image, u is the
interpolation kernel, and s is the distance between the pixel of interest and its neighboring
pixels. As noted in [15], the accuracy and efficiency of the algorithm lies solely on the
interpolation kernel, u. It was empirically found that the smallest possible scale factor
without compromising the detection accuracy of the 10 input images is 0.25. All the faces
in the original image were detected with no false positives. The performance report after
applying bicubic down-sampling is given in Table 2.
Table 2. Performance report after applying bicubic down-sampling.
Section
Imresize
CalculateIntAndSqIntImages
FaceDetection
Average Time Per Image

Total Time
(sec)
84.74
0.16
1050.27
0.19
49.14
1000.78

Occurance
10
10
10
10
90
315,100

Average
Time (sec)
8.47
0.02
105.03
0.02
0.55
0.003
113.52

Percentage of
Total Time (%)
7.46
0.01
92.52
0.02
4.33
88.16

The down-sampling calculations are carried out in the Imresize function. As seen in the
performance report, the size reduction of the input image has a positive performance affect
on all functions that carry out operations on the image. On average, a speedup factor of
16.5 is achieved after down-sampling the input image to a quarter of its original size.
An additional optimization applied was to combine the CalculateIntegralImage
and CalculateSqIntegralImage functions together. This is a logical step given that
these functions have a similar structure and operate on the same data. The resulting
function name is CalculateIntAndSqIntImage, and completes execution in
approximately half the time originally required to calculate the integral and square integral
images separately. The reduction in time is a result of not having to fetch the same image
data from extended memory (SDRAM) twice.

3. Optimization Using Custom Instructions
Configurable custom processors are becoming an ever more popular implementation
technology of choice for addressing the demands of complex embedded applications.
Unlike traditional hardwired processors that consist of a fixed instruction set from which
application code is mapped; configurable processors can be augmented with application
specific instructions, implemented as hardware logic to optimize bottlenecks. This lends
towards a method for hardware-software partitioning whereby the efficiency of hardware
and the flexibility of software are integrated.


7

There are a number of benefits in extending a configurable processor with custom
instructions. First, transparency; the added custom instructions will improve the
performance of the tasks for which they are designed with minor changes to the original
code. Second, rapid development; there is a wide variety of off-the-self configurable cores
that could be used as a base for development. Additional instructions could be integrated
into the processor core as the need to extend its computational capabilities arises. Finally,
low-cost access to domain specific processors; generally the fundamental characteristics of
an application area is similar. These characteristics can be summarized as a set of
instructions and applied to a variety of similar applications, for example, multimedia
applications [16].
Unfortunately, there are two minor drawbacks to using custom instructions. For one,
additional hardware is required, although this is becoming less of an issue as embedded
technologies become more economical. Secondly, as the custom instructions are directly
integrated into the processor’s pipeline, the maximum operating frequency may be
degraded if the instruction is poorly designed. Adding custom instructions is a proven
optimization technique that has been applied to a wide range of embedded applications.
Some published examples include, embedded real-time operating systems (RTOS) [17],
biometrics [18], and multimedia [19].
3.1 Custom Instruction Design Flow
The design flow for identifying and integrating custom instructions into configurable
processors is summarized in Figure 1. This is a generic framework that could be applied
to any application. Firstly, the software code is profiled to reveal bottlenecks that could be
alleviated with the introduction of custom instructions. Once the hardware module for the
instruction is implemented and tested, it is added to the processor and the whole system is
regenerated. Then the software code is updated to make use of the new instructions.
Finally, the functionality of the system is verified to ensure bugs are not introduced with
the new instruction. This process is repeated until either the performance requirements or
resource limits are met.
3.2 Extending Nios II with Custom Instructions
All of the Altera Nios II processor cores are designed to support up to 256 new custom
instructions. The logic for the new instructions is directly connected to the arithmetic logic
unit (ALU), as illustrated in Figure 2. In the face detection application, this comes does to
applying the custom instructions to the Viola-Jones algorithm The Viola-Jones application
code which includes the initial software optimizations is profiled using performance
counters. As expected, the FaceDetection function, more specifically, the
RunClassifierCascade function is the most time consuming. It should be noted that
it is not the RunClassifierCascade function itself that is time consuming, but due to
the large number of times it is called, the accumulated time is large. Hence, the overall
processing time could be improved if either the number of calls to
RunClassifierCascade or its execution time is reduced. Since reducing the number
of calls to RunClassifierCascade is going to compromise the accuracy of the face
detector, the focus is placed on reducing the execution time of the function itself. Any time
savings made in each function call will correspond to a large overall saving.
The main operations carried out in the RunClassifierCascade function are
addition and multiplication of integer values associated with the calculation of indexes
into the input image. There are also floating point multiply and addition operations that are
required for image normalization and weighting of the classifier features. Since all the

8


integer related instructions are already implemented in hardware, it is believed that there
are opportunities to improve the performance of the floating point operations through
custom instructions.

Fig. 1. Custom instructions design flow.

Fig. 2. Connection of custom instruction logic
with the Nios II ALU.

Floating point multiply is the first operation of interest as it is used most frequently.
Further profiling of the RunClassifierCascade function indicates that this operation
takes up roughly 559 seconds of the total 1135 seconds, approximately 49% of the total
processing time. By default all floating point arithmetic are emulated in software. For
these reasons the first operation to be implemented as a custom instruction is the floating
point multiply. The next instruction that was implemented was floating point add/subtract
Instructions. A summary of the performance reported for the face detector after the
addition of the floating point multiply instruction and add/subtract Instruction are
presented in Table 3 and Table 4 respectively.
Table 3. Performance with the use of the floating point multiply custom instruction.
Section
Imresize
FaceDetection

Total Time
(sec)
56.5
0.16
571.17
0.19
36.62
534.03

Occurance
10
10
10
10
90
315,100

Average
Time (sec)
5.65
0.02
57.12
0.02
0.41
0.002
62.79

Percentage of
Total Time (%)
9
0.03
90.98
0.03
5.83
85.06

Table 4. Performance with the use of the floating point addition/subtraction custom instruction.
Section
Imresize
FaceDetection

Total Time
(sec)
57.46
0.16
393.61
0.19
34.78
358.43

Occurance
10
10
10
10
90
315,100

Average
Time (sec)
5.75
0.02
39.36
0.02
0.39
0.001
45.13

Percentage of
Total Time (%)
12.73
0.04
87.23
0.04
7.71
79.43


9

3.4 Optimization of the Imresize Function
The next most time consuming function to focus on is Imresize. Since the bicubic
resizing algorithm used in the Imresize function is not specific to the Viola-Jones
algorithm and could also be applied to many other applications, it is an ideal candidate for
optimization. As it stands, the Imresize function accounts for 13% of the total time.
By examining the bicubic interpolation algorithm more closely, it becomes evident that
the coefficients calculated by the interpolation kernel can be fixed constants. More
importantly, the exact value of these constants is not essential [20]. A graphical illustration
for computing an interpolated value is depicted in Figure 3. In essence, the two
dimensional convolution like calculations described in Equation 1 can be decomposed to
five sets of one dimensional operations. The coefficients, a0, a1, a2, and a3, as seen in
Figure 3 all add to one; hence, the intermediate sums will never exceed the largest pixel
value of 255. Also, the intermediate results after multiplication are truncated to 8-bit
integer values.

Fig. 3. Graphical description for calculating an interpolation value.

A series of experiments were conducted using the original implementation of the bicubic
interpolation algorithm to look at the behavior of a0, a1, a2 and a3 for a variety of
different scale factors. The following observations were made:
• a0 and a3 have the same values
• a2 is roughly four times larger than a0 and a3 while a1 is roughly two times
larger than a0 and a3
• when the coefficients are set to the same value and are close to zero, the resulting
image is very dark (basically black)
• when the coefficients are set to the same value and close to one, the resulting
image is extremely noisy
Based on these observations, the chosen values for a0, a1, a2, and a3 are 0.125, 0.5, 0.25,
and 0.125, respectively. These coefficients are chosen because they correspond to dividing
the pixel values by 2, 4, and 8, that is, shifting of the pixel values to the left by 1, 2, and 3
bits, respectively. The images produced based on the new implementation of the
Imresize function using these coefficients is identical to the original implementation
and that from Matlab — Matlab’s implementation is used as an additional confirmation
step. The Imresize function that utilizes the coefficients, a0, a1, a2, and a3 is referred to
as the “new implementation”, while the implementation that makes use of the bicubic
kernel is referred to as the “original implementation”.

10


(a)

(b)

(c)

Fig. 4. Examples of images resized to a quarter of their original size. The methods used are: (a)
original implementation, (b) new implementation, and (c) Matlab’s bicubic resize function.

The performance of the original and new implementation of the Imresize function is
summarized in Table 5; the values obtained are averaged across all 10 face images. The
new implementation is 115 times faster than the original implementation.
Table 5. Performance summary of the original and new implementation of the Imresize.
Section
Original Imresize
New Imresize

Average Clock Cycles
459,693,061
3,949,404

Average Time (sec)
5.75
0.05

4. The Effect of Data and Instruction Caches on Performance
It is also important to investigate what the effects data and instruction caching behavior
and size have on performance. Initially, the Nios II processor is configured with the
default data cache settings, that is, an on-chip memory size of 16 KB with a data cache
line size of 4 bytes. According to the Nios II core documentations, if the line size is greater
than 4 bytes, data retrieval from extended memory (SDRAM in our case) is pipelined;
hence reducing the impact of data transfer latency. A new system with exactly the same
configuration but with the data cache line size increased to 32 bytes is generated. When
the face detector and face detector beta programs are ran on this system, the total
execution times are 466.23 and 394.18 seconds, respectively – the difference between the
two execution times are less. These results give a positive indication that the processor’s
caches have an influence on performance.
The next experiment is to look at the effects of altering the size of the data cache. A
series of eight systems with all the possible data cache sizes are generated. All these
systems have a data cache line size of 32 bytes and the instruction cache size is fixed to 4
KB. Table 6 summarizes these results. The usage of other resources such as DSP blocks,
PLLs, and pins remain the same. There are minor fluctuations in the amount of LEs used
and operating frequencies, but this is likely due to the variability in optimizations by the
synthesis and fitting tools. Logically, the total amount of memory bits utilized linearly
scales with the data cache size. As seen in Table 6, the size of the data cache does have an
affect on the performance of the programs, particularly in the size range from 0.5 to 16
KB. Also, the performance continues to improve with a larger data cache.
A similar system but with data and instruction cache sizes of 64 KB (the largest cache
sizes possible) were also generated on a Stratix EP1S40 development board (with roughly
four times more resources than EP1S10) to confirm that no further improvements were
possible with an instruction cache larger than 16 KB. Similarly other computational
functions including Divide, Compare, and Round were added as custom instructors and
similar speed-ups were observed.


11

Table 6. Summary of the resource usage and executions times for varying data cache sizes.
Data Cache Size
(KB)
0.5
1
2
4
8
16
32
64

LE (%)
66
68
66
66
66
68
68
68

Total Memory
Bits (%)
5.65
6.13
7.07
8.96
12.71
20.19
35.1
64.8

Face Detector
(sec)
589.26
555.3
527.24
507.48
481.64
466.23
458.8
455.79

Face Detector
Beta (sec)
483.07
452.11
433.59
421.03
406.13
394.18
389.05
386.89

5. Conclusion
This paper investigated the effects of replacing software bottleneck operations of a FaceDetection System based on Viola-Jones algorithm with custom instructions on
performance. Table 7 presents a summary of the new instructions implemented along with
a measure of their efficiency — in-order for comparisons to be made fairly, the floating
point multiply custom instruction is re-synthesized without the use of DSP blocks.
Table 7. Speedup, resource usage, and efficiency measure for each custom instruction.
Floating Point Operation
Multiply
Add/Sub
Divide
Compare
Round

Resource (LE) Speedup
1,019
18
806
21
1,061
11
77
16
354
37

Speedup/Area (10-3)
18
26
10
208
105

These results indicate that the floating point compare custom instruction is by far the most
efficient in terms of speedup to area, even though it has a low overall speedup factor when
integrated with the face detector application. On the other hand, even though the floating
point multiply instruction has a reasonably low speedup to area ratio, when used in the
face detector application the speedup for this instruction is high, in part because it is one
of the most commonly used operations.
As the Viola-Jones face detection algorithm is primarily dominated by control
operations and calculations involving 32-bit integer or floating point numbers, very little
benefit is likely to result from the movement of larger functions to hardware.
An inadvertent result revealed through this investigation is that both the size and
behavior of the caches, specifically the instruction cache, has a significant affect on the
software performance. Experiments have shown that the total execution time may
noticeably fluctuate depending on the code or instruction cache size. The implication of
this result is that, it is difficult to determine the effectiveness of the optimizations applied
— even with custom instructions; since changes to the software code results in a change to
the code size and hence caching behavior. Lastly, it has been shown that incremental
changes to the software code can add up to substantial reductions in the total execution
time. However, the extent and effectiveness of these optimizations is largely attributed to
the designer’s experience.
Acknowledgement
This project was partly supported by a grant from the Australian Government Department
of the Prime Minister and Cabinet. NICTA is funded by the Australian Government's

12


Backing Australia's Ability initiative and the Queensland Government, in part through the
Australian Research Council. However the majority of work was done at The University
of Auckland in New Zealand.
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.

M. H. Yang, D. J. Kriegman, and N. Ahuja: Detecting faces in images: a survey. In IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 24 (2002) 34-58
E. Hjelmas and B. K. Low: Face detection: a survey. In Computer Vision and Image
Understanding, vol. 83, (2001) 236-274
R. W. Frischholz and U. Dieckmann: BiolD: a multimodal biometric identification system. In
Computer, vol. 33 (2000) 64-68
D. P. Huttenlocher, G. A. Klanderman, and W. J. Rucklidge: Comparing images using the
Hausdorff distance. In IEEE Transactions on Pattern Analysis and Ma-chine Intelligence, vol.
15 (1993) 850-863
Ting Shan, Brian C. Lovell, Shaokang, Chen and Abbas Bigdeli: Reliable Face Recognition for
Intelligent CCTV. In Proc. of Safeguarding Australia 2006- The 5th Homeland Security
Summit & Exposition (2006) 356-364
T.-K. Kim, S.-U. Lee, J.-H. Lee, S.-C. Kee, and S.-R. Kim: Integrated approach of multiple face
detection for video surveillance. In Proc. of Int. Conf. on Pattern Recognition, vol. 2 (2002)
394-397
T. Theocharides, G. Link, N. Vijaykrishnan, M. J. Irwin, and W. Wolf: Embedded hardware
face detection. In Proc. of the 17th Int. Conf. on VLSI Design (2004) 133-138
H. A. Rowley, S. Baluja, and T. Kanade: Neural network-based face detection. In IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 20 (1998) 23-38
H. A. Rowley, S. Baluja, and T. Kanade: Rotation invariant neural network-based face
detection. In IEEE Computer Society Conf. on Computer Vision and Pattern Recognition
(1998) 38-44
B. D. T. Inc.: Using General-Purpose Processors for Signal Processing. In ARM Developers'
Conf. (2004)
R. McCready: Real-Time Face Detection on a Configurable Hardware System. In Proc. of The
Roadmap to Reconfigurable Computing, 10th International Workshop on Field-Programmable
Logic and Applications (2000) 157-162
M. S. Sadri, N. Shams, M. Rahmaty, I. Hosseini, R. Changiz, S. Mortazavian, S. Kheradmand,
and R. Jafari: An FPGA Based Fast Face Detector. In Global Signal Processing Expo and Conf.
(2004)
D. M. Lewis, D. R. Galloway, M. Van Ierssel, J. Rose, and P. Chow: The Transmogrifier-2: a 1
million gate rapid-prototyping system. In IEEE Transactions on Very Large Scale Integration
(VLSI) Systems, vol. 6 (1997) 188-198
R. Lienhart, A. Kuranov, and V. Pisarevsky: Empirical Analysis of Detection Cascades of
Boosted Classifiers for Rapid Object Detection. In DAGM, 25th Pattern Recognition
Symposium (2003) 297-304
R. Keys. Cubic convolution interpolation for digital image processing. In IEEE Transactions on
Acoustics, Speech, and Signal Processing, vol. 29 (1981) 1153-1160
A. Bigdeli, M. Biglari-Abhari, S. H. S. Leung, and K. I. K. Wang: Multimedia extensions for a
reconfigurable processor. In Proc. of 2004 International Symposium on Intelligent Multimedia,
Video and Speech Processing, (2004) 426-429
T. F. Oliver, S. Mohammed, N. M. Krishna, and D. L. Maskell: Accelerating an embedded
RTOS in a SoPC platform. In Proc. of TENCON Conference, vol. 4 (2004) 415-418
H. Tsutsui, T. Masuzaki, T. Izumi, T. Onoye, and Y. Nakamura: High speed JPEG2000 encoder
by configurable processor. In Proc. of Asia-Pacific Conf. on Circuits and Systems, vol. 1 (2002)
45-50
Z. GuangWei and L. Xiang: An efficient approach to custom instruction set generation. In IEEE
Int. Conf. on Embedded and Real-Time Computing Systems and Applications (2005) 547-550
W. H. Press, W. T. Vetterling, S. A. Teukolsky, and B. P. Flannery: Numerical recipes in C: the
art of scientific computing. Second ed. Cambridge University Press, New York (1992)

Face detection on_embedded_systems

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Face detection on_embedded_systems (20)

More from r_sadoun (6)

Recently uploaded (20)

Face detection on_embedded_systems