EMBEDDED CONVOLUTIONAL FACE FINDER
Sébastien Roux, Franck Mamalet and Christophe Garcia
France Telecom division R&D, 28 Chemin du Vieux Chêne, 38243 Meylan, France
e-mail: <first name>.<last name>@francetelecom.com
In this paper, a high-level optimization methodology is applied
for the implementation of the well-known Convolutional Face
Finder (CFF) algorithm for real-time applications on cellular
phone, such as teleconferencing, advanced user interfaces,
pictures indexing and security access control. This face detector is
based on a feature extraction and classification technique which
consists in a pipeline of convolutions and subsampling operations.
Design of embedded systems must find a good trade off
between performance and code size due to the limited amount of
resource available. We propose a methodology to cope with the
main drawbacks of the CFF original implementation like floating-
point computation and memory allocation, to allow parallelism
exploitation and perform algorithm optimizations. Results show
that our embedded face detection system can accurately locate
faces with less computational load and memory cost. It runs on a
275MHz Starcore DSP at 9 QCIF images/s with state-of-the-art
detection rates and very low false alarm rates.
When embedding new services on mobile devices, one of the
strongest constraints is the limited computational resources. Low
memory capacities, low CPU frequency and lack of specialized
hardware like floating point unit are some of the major differences
between a PC and an embedded platform. Unfortunately, advanced
algorithms are usually developed on PC without any
implementation restriction in mind. Thus, porting application on
power-constrained embedded systems is a challenging task and
requires strong algorithmic, memory and software optimizations.
Advanced user interface, security access control, model based
video coding, image and video indexing are some of the
applications that rely on face detection. In recent years, numerous
approaches for face detection have been proposed. A survey was
published by Yang et al.  in 2002. In this paper, face detection
techniques are classified in three main categories:
feature invariant approaches ,
template matching methods ,
appearance-based methods .
A recent technique that belongs to the third category, called
Convolutional Face Finder (CFF) has been introduced by Garcia
and Delakis  which leads to the best performance on standard
face databases. The CFF is an image-based neural network
approach that allows robust detection, in real world images, of
multiple semi-frontal faces of variable size and appearance, rotated
up to +/- 20 degrees in image plane and turned up to +/- 60
Addressing both face
implementation on embedded system has been considered in recent
years by Tang et al.  for cascade adaboost classifiers  on
ARM based mobile phones. The Adaboost technique was also used
in  for implementing a hybrid face detector on a TI DSP.
Another way to achieve resource constrained implementation is to
design dedicated hardware for face detection. In , the authors
proposed an ASIC implementation of the face detector introduced
by Rowley et al .
However, in real time embedded implementations one often
has to trade off among high detection rates, fast run time and small
code size. In most cases, the side effect of embedding a face
detector is the reduction of the algorithm efficiency. We will show
that we have achieved both efficiency and speed objectives (10
images/s) with our CFF implementation.
The remainder of the paper is organized as follows. An
overview of the Convolutional Face Finder technique is given in
Section 2. Section 3 presents the methodology used for embedding
such an algorithm. Section 4 details this methodology on the CFF
case study. Experimental results for DSP (Starcore SC140) and
RISC (XScale) based platforms are provided in section 4. Finally,
conclusions and perspectives are drawn in section 5.
detection performance and
2. CFF ALGORITHM OVERVIEW
Fig.1. Convolutional Face Finder pipeline.
The Convolutional Face Finder was presented in  and relies on
Convolutional Neural Networks introduced and successfully used
by LeCun and al. . It consists in a pipeline of convolutions and
subsampling operations (Fig. 1). This pipeline performs automatic
2851424403677/06/$20.00 ©2006 IEEEICME 2006
feature extraction in image areas of size 32x36, and classification
of the extracted features, in a single integrated scheme. In , the
authors present both the training methodology to learn the
coefficients using back propagation, and the face localization
process when training has been completed. In this paper we will
only consider the face localization process. Fig. 2 presents in detail
the steps of this face localization process:
• A coarse detection is first performed as follows. The CFF is
applied on a pyramid of scaled versions of the original image
(Fig. 2-1) in order to handle faces of different sizes: each scale
produces a map of faces candidates (Fig. 2-2) which is fused
back to the input image resolution and produce clusters of
positives answers (Fig. 2-3). For each cluster a representative
face is computed as the centroid of its candidate face centers and
sizes weighted by their individual network responses (Fig. 2-4).
• Then a fine detection takes those candidates as input and applies
locally the CFF on a small pyramid around the face candidate
center position. The volume of positive answers is considered to
take the classification decision of face or non-face (Fig. 2-5).
Finally, overlapping candidates are fused to remove multi
detections of the same face.
Fig. 2. The different steps of the process of face localization.
3. PORTING CFF TO EMBEDDED PLATFORMS:
MAIN ISSUES AND METHODOLOGY
In order to implement complex algorithm on embedded target
processor, compilers are the tools to optimize the instructions flow.
In the last decade, many research activities have been carried out in
instruction-flow optimizations  and optimizing compilers ,
and some have led to industrial products such as the Metrowerks
compiler for SC140 . However, compilers can only cope with
the instructions flow optimization and parallelization.
Even if these compilers avoid mostly any human assembly
programming, many optimizations have to be done by manual
Our approach is based on iterations of high-level code
optimizations and profiling to focus first on the most CPU resource
consuming functions. When dealing with an algorithm such as
CFF, the first step towards an embedded implementation is to
avoid floating point calculation. This step is done thanks to a
fractional transformation in accordance with data dynamics and
processor data paths. This also requires a strong verification of the
accuracy of these transformations which can otherwise lead to
incorrect results. The next steps of the methodology are iterations
of a tri-optimization flow (code, memory and algorithm) controlled
by an on-target profiling (fig. 3).
Profiling tools depend on the target platform: for instance, we
use the VTune software on Xscale based platform to profile the
compiled code directly on target, and global timing information to
evaluate the speed up factor after each optimization iteration.
Fig. 3. Diagram of followed methodology.
We will illustrate our methodology on the CFF
implementation, which starting point was a floating point
arithmetic version and required a memory allocation of 3.8 MBytes
to process a QCIF format image (176x144 pixels). The reference
complexity analysis of the floating point version of the CFF shows
that it requires 3s to compute a single QCIF image on a 624MHz
Xscale processor. Hereafter, we present in detail each step of this
methodology and the achieved performance results.
4. OPTIMIZING THE CFF ALGORITHM
4.1. Fractional transformation
The reference software of the CFF was entirely written using
floating point arithmetic. Mobile embedded target platforms lack
floating-point hardware accelerator for power consumption
reasons. Floating-point computations are usually implemented by
software, but these are high CPU consuming functions. The first
step towards embedding the algorithm is to transform the floating
point computations into fractional ones. Since one of our target
platforms was the 16 bits DSP Starcore SC140, fractional Q15
arithmetic  was required (Q31 arithmetic may be used when
more precision is needed).
The main advantage of the CFF algorithm is that the results of
the subsampling layers S1 and S2 pass through a hyperbolic
tangent function, thus reducing the risk for common issues of fixed
point computations such as arithmetic dynamic expansion and
saturation. A simple methodology was used to normalize and
transform each stage coefficient in fixed-point arithmetic and
compare the results with the floating point version.
The main constraint of this transformation was to keep the
efficiency of the face detector. The benchmarking was done on
different test sets of images, including the CMU Test Set (the most
widely used data set in the literature). Table 1 gives the detection
rates of the floating and fixed point versions for different
configurations of the CFF (varying output threshold and minimum
faces detection size).
TABLE 1: results of CFF on different test sets for the floating
and fixed point versions
The comparison of the floating point and fixed point versions
shows up no significant loss in efficiency, and detection rates are
equivalent to the previously published ones in . They are even
slightly better on part of the selected test sets. What is especially
noticeable about CFF efficiency is the low level of false alarms and
even after the fractional transformation.
4.2. Memory optimization
Due to the computational redundancy in the CFF algorithm, the
reference software was processing layer by layer on the whole
image (or scaled versions of the original image). This configuration
is not suitable for an embedded platform since even for small QCIF
images, 3.8 MBytes were allocated (for instance, the targeted
SC140 DSP platform embeds only 512kB of SRam).
In order to reduce this memory allocation without increasing
the required amount of computations, a study was made on the data
dependency in the algorithm. Fig. 4a shows the amount of data
needed in each layer in order to compute a single output of each
neuron layer N1. This figure is similar to Fig.1 restricted to one
feature map by layer.
Fig. 4b illustrates the differential computation between two
neighbouring outputs (south side) of neuron layer N1. Slashed
(resp. unslashed) grey parts are unused (resp re-used) previously
computed data, whereas dark rectangle are newly computed data.
Since Fig. 4b shows that intermediate computation from
previous line has to be kept as input of the layer C2 and N1, the
maximum gain in terms of memory footprint is achieved for a line
by line processing of the output of N1 layer. Thus, in the final
implementation, in order to compute one output line of the layer
N1, we use 7 input lines of this layer. These input lines can be
computed line by line in layer S2 using two output lines of layer
C2. These two output lines require four input lines for the layer
C2. Two of these four output lines are common with the previously
computed lines, and the two others require four output lines of the
layer C1. To end with, these four output lines are computed using
eight lines of the input image.
Fig. 4. CFF data flow. a) amount of data needed in each layer,
b) differential computation between two neighbouring outputs
CFF algorithm analysis for the full image processing (resp. the
line by line processing) shows that memory allocation is about
10.25*W*H+… (resp. 66*W+…), W and H being the width and
height of the input image.
For a QCIF image the gain in memory footprint is about 21.
Other memory allocation optimizations (e.g. on scaled image
computation) have been made on the reference software leading to
a memory footprint of 220kB compared to the 3.8 Mbytes of the
4.3. Code optimization: parallelism exploitation
One of our target embedded platforms is a Starcore SC140 DSP
which has 4 ALUs and Multiplier capabilities. This processor is
able to load eight 16 bits-words and to compute 4 Multiplication-
accumulations (MACs) in one cycle. The main limitation to take
advantage of this parallelism is that one needs to satisfy data's
alignment constraints: the Move.4F instruction which loads four
16bits-word data is only allowed for an eight bytes aligned pointer
and can be generated automatically by the compiler by appropriate
C code re-writing and alignment directive use.
Let us analyse the first layer (C1) which is pointed out by the
profiling tool as the most complex step of the CFF algorithm: each
of the four feature maps of this layer consists in a convolution by a
5x5 kernel. Without any parallelization one convolution requires
25 data loads, 25 coefficient loads, 25 MACs instructions and one
store instruction. Since the Starcore is able to compute four MACs
in one cycle, the theoretical minimum cycle count for processing
25 MACs (without load and store count) is [25/4] = 7 cycles. The
Starcore is able to process two (single or multiple) load
instructions by cycle (in parallel with the MACs instructions).
Thus, without aligned loads instructions, one convolution would
require at least [(25+25+1)/2] = 26 cycles. So, the main goal to
optimize such a function is to reduce the number of load and store
instructions by using the Move.4F instruction.
N1 S2 C2 S1 C1
Faces size 36 to 300 pixels high 18 to 300
Threshold 10 17 17
CMU 84,896 80,120 87,992
CINEMA*87,328 82,971 82,97 4
WEB* 87,982 83,97 0 91,982
CMU 86,754 81,370 88,203
CINEMA*88,416 82,253 85,149
WEB* 88,981 86,171 92,385
* CINEMA and WEB are test sets of respectively 276 and 499 faces
kindly provided by C.Garcia 
However, the 5x5 convolution processing is done on any image of Download full-text
the pyramid whose width is not necessarily multiple of 4. Thus if
the first top-left pixel in the image is 8 bytes aligned, the first pixel
on the second line will probably not be aligned preventing from
any use of multiple load instruction on these data.
The proposed solution in order to reduce the number of load
instructions per convolution consists in factorizing the coefficients
loads for several processing of the 5x5 convolution (multi-sample
Convolutions are done by 25 iterations on the whole block of
pixels. At each iteration, groups of four multiplication
accumulations with a single coefficient are done. This requires a
temporal store and load of intermediate processing, but, since this
intermediate matrix can be 8 bytes aligned, four intermediate
computations can be loaded or saved in a single instruction. As a
result, the amount of load and store instructions needed for this
modified version is 25+37.5*S compared to the 51*S instructions
required for the initial version (where S=(W-4)*(H-4)).
So, when processing four output line of the layer C1 as
depicted in the previous paragraph, the gain in terms of load/store
instructions is 26.4 % for a QCIF image (W = 176, H = 8).
Since the SC140 compiler achieves the best instruction flow
parallelization, we get the same gain in term of number of cycles
by convolution with this factorized version.
This optimization may also be applied on processors using
SIMD instructions such as WMMX instructions on Xscale
embedded processor. The efficiency of this optimization on these
processors has not been evaluated yet.
4.4. Performance results
Table 2 summarizes the speed up factor obtained on a QCIF video
test sequence (120 first frames of the Mpeg Foreman sequence)
after several others iterations of the optimization methodology.
Furthermore, as depicted before, the memory footprint has
been reduced from 3.8 MBytes to 220 kBytes by the memory
TABLE 2: CFF processing speed
5. CONCLUSION AND PERSPECTIVES
In this paper, we have presented the implementation of a state-of-
the-art face detector on two kinds of programmable embedded
platforms. We have shown that both high detection rates and fast
processing are achieved by applying our optimization flow
methodology. Memory and code restructuring in conjunction with
algorithm adaptation lead to significant improvement. Indeed, we
obtain a speed-up factor of 14 on an Xscale PXA27x based
platform, and video processing at 9 QCIF fr/s on a Starcore DSP.
Efficiency is maintained high, with detection rate of 87 % on the
CMU test set and only 4 false alarms.
One of our final objectives is to provide an embedded face
recognition system for biometrics applications. Usually, face-based
identification systems need precise face detection but also fine
facial feature localization. The first step depicted in this paper was
the real time implementation of this face detector by software
optimizations. The second step is to detect facial features, and we
are now working on the implementation of a facial feature position
extractor based on the same principles which is called C3F for
Convolutional Face Feature Finder .
Furthermore, this study points out that the pipeline of
convolutional and subsampling filters denotes high intrinsic and
hidden parallelisms which will be exploited in future works with
dedicated hardware implementation of CFF and C3F.
 M. Yang, D. Kriegman, and N. Ahuja, “Detecting Faces in
Images: A Survey,” IEEE Trans. Pattern Analysis and Machine
Intelligence, vol. 24, no. 1, pp. 34-58, Jan. 2002.
 K.C. Yow and R. Cipolla, “Feature-Based Human Face
Detection,” Image and Vision Computing, vol. 15, no. 9, pp. 713-
 I. Craw, D. Tock, and A. Bennett, “Finding Face Features,”
Proc. Second European Conf. Computer Vision, pp. 92-96, 1992.
 K.-K. Sung and T. Poggio, “Example-Based Learning for
View-Based Human Face Detection,” IEEE Trans. Pattern
Analysis and Machine Intelligence, vol. 20, no. 1, pp. 39-51, Jan.
 C. Garcia and M. Delakis, “Convolutional Face Finder: A
Neural Architecture for Fast and Robust Face Detection,” IEEE
Trans. Pattern Analysis and Machine Intelligence, vol. 26, no. 11,
pp. 1408-1423, Nov. 2004.
 X. Tang, Z. Ou, T. Su and P. Zhao, “Cascade AdaBoost
Classifiers with Stage Features Optimization for Cellular Phone
Embedded Face Detection System,”. ICNC (3) 2005, pp. 688-697.
 P. Viola and M. Jones, “Rapid Object Detection Using a
Boosted Cascade of Simple Features,” Proc. Int’l Conf. Computer
Vision and Pattern Recognition, vol. 1, pp. 511-518, 2001.
 J.-B. Kim, Y. H. Sung, S.-C. Kee, “A Fast and Robust Face
Detection based on Module Switching Network, ” Int. Conf. on
Automatic Face and Gesture Recognition, pp 409-414, May 2004.
 T. Theocharides, et al, “Embedded Hardware Face Detection”,
Int. Conf. on VLSI Design, pp. 133-137, Jan. 2004.
 H. Rowley, S. Baluja, and T. Kanade, “Neural Network-Based
Face Detection,” IEEE Trans. Pattern Analysis and Machine
Intelligence, vol. 20, no. 1, pp. 23-38, Jan. 1998.
 Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-
Based Learning Applied to Document Recognition,” Proc. IEEE,
vol. 86, no. 11, pp. 2278-2324, 1998.
 V. Tiwari, S. Malik, and A.Wolfe, “Compilation techniques
for low energy: An overview,” Proc. of the Symposium on Low
Power Electronics, 1994.
 S. Muchnick, “Advanced
Implementation.,” Morgan Kauufmann Publishers, 1997.
 V. Palanciuc, D. Bade, E. Flamand, C. Ilasa, “Spill Code
Reduction Technique for EPIC architectures application in the
Metrowerks StarCore C compiler ,” Workshop on EPIC, 2001.
 A. Bateman and I. Paterson-Stephens, The DSP Handbook,
Algorithms, Applications and Design Techniques, Prentice Hall,
pp. 128-130, 2002.
 S. Duffner, C. Garcia, “A Connexionist Approach for Robust
and Precise Facial Feature Detection in. Complex Scenes,” Int.
Symp. on Image and Signal Processing and Analysis (ISPA), 2005.
Compiler Design &
Fixed point and
0.3 fr/s - 10 fr/s
4.5 fr/s 9 fr/s 32 fr/s