EFFICIENT REALTIME FPGA IMPLEMENTATION OF THE TRACE TRANSFORM
Suhaib A. Fahmy§ , Christos-Savvas Bouganis§ , Peter Y.K. Cheung§ and Wayne Luk‡
§ Department of Electrical and Electronic Engineering
Imperial College London, Exhibition Road, London, SW7 2BT, England.
email: {s.fahmy, ccb98, p.cheung}@imperial.ac.uk
‡ Department of Computing
Imperial College London, 180 Queen’s Gate, London, SW7 2AZ, England
email:wl@doc.ic.ac.uk
ABSTRACT
The Trace Transform is a novel image transform that is able
to exhibit useful properties such as scale and rotation invari-
ance and occlusion robustness. As a result, it is particularly
suited to a variety of classification and recognition tasks in-
cluding image database search, token registration, activity
monitoring, character recognition and face authentication.
The main obstacle to the widespread use of the transform
is its high computational complexity. This has precluded
a detailed investigation of transform parameters. This pa-
per presents an architecture and implementation of a Trace
Transform engine on a Virtex-II FPGA. By exploiting the in-
herent parallelism in the algorithm and the use of optimised
functional blocks, a huge performance gain is achieved, ex-
ceeding realtime video processing requirements for a 256 x
256 image.
1. INTRODUCTION
The study of image recognition relies heavily on properties
in 2D shape and texture. Indeed, the extraction of features
from an image has been used extensively in image classi-
fication and matching. However, many of these feature ex-
traction methods focus primarily on properties seen from the
human perspective. It is, though, useful to expand the hori-
zon, using features which may not have meaning to humans,
but which perform well in characterising complex images.
With this in mind, Kadyrov and Petrou proposed the Trace
Transform in [1] and developed it further in [2, 3, 4]. The
transform is a redundant representation of an image, from
which features can be extracted. It is a generic transform in
the sense that the mathematical definition is extensible.
This work was partially funded by the UK Research Council under
the Basic Technology Research Programme (GR/R87642/02) and by the
EPSRC Research Grant (EP/C549481/1).
The strength of the transform lies in its ability to extract
features that are robust to affine transformations, occlusion
and even non-linear deformations. In fact, careful selec-
tion of the transform functionals [2] can allow not just ro-
bustness, but recovery of transformation coefficients, which
would allow an affine-transformed image to be returned to
its original state.
The Trace Transform has proved to be a very power-
ful tool, showing excellent results in image database search,
industrial token registration, activity monitoring, character
recognition and face authentication [5, 6]. The primary ob-
stacle to further application investigation has been the com-
putational complexity of the algorithm.
FPGAs provide an ideal platform for accelerating the
Trace Transform through exploitation of the inherent paral-
lelism existent within the algorithm. Furthermore, the flex-
ibility of the platform suits the generic nature of the trans-
form, since alternative functionals can be swapped in an out
with ease. Acceleration of the Trace Transform allows for
further investigation of the transform in terms of both func-
tionals and applications.
We present in this paper the first hardware implementa-
tion of the Trace Transform. Our architecture is:
• fast: with performance exceeding real-time process-
ing requirements;
• simple: occupying a very small area on the FPGA;
• flexible: allowing us to investigate Trace functionals
thoroughly.
2. ALGORITHM
2.1. Overview
It is important before considering our implementation, that
we introduce the algorithm briefly so as to clarify the prin-
ciples that will be discussed in this paper. The Trace Trans-
form of an image is a transformation from the spatial domain
1-4244-0 312-X/06/$20.00 c©2006 IEEE.