In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds). IAPRS. Vol. XXXVIII. Part ЗА - Saint-Mandé, France. September 1-3. 2010 
291 
perfectly discriminates road signs from noise compared to simple 
gradient-based score. 
Hybrid algorithm 
Figure 5: ROC curve for our algorithm applied to database CD3 
and using two kinds of score calculation 
To illustrate the interest of primitives fusion, we used the same 
image database which we processed using different values of the 
parameter e re f- This parameter changes the influence of prim 
itives with respect to the error e e . The lower the value of e re f 
(closer to zero), the later the appearance of edge primitive influ 
ence in the convergence process. This means that for low values 
of e re f, edge primitive effects lose their influence and the preci 
sion of convergence should be lower. On the other hand, if we 
use higher values for this parameter, we lose in efficiency and 
an individual would require more iterations to match an object. 
This phenomenon is observed in figure 6 which shows the mean 
number of iterations required (generations) for the best individual 
to converge (minimum error) on a road sign, and the mean error 
that the best individual reaches. We note that if we choose a large 
value for e re f, the mean of number of generations is about 3 times 
higher. We therefore choose a value of e re f around the point of 
the convergence point of the curves. The algorithm needs around 
15 iterations with 20 individuals to achieve a very precise conver 
gence compared to the needed 50 iterations with 100 individuals 
for (De La Escalera et al.. 2004). In order to show the precision 
Figure 6: Mean of number of generation and mean of final error 
of best idividual w.r.t e re f 
of the convergence, figure 7 deals with different views of sam 
ples which come from original images. The fine-detection algo 
rithm found an ideal couple {template, configuration} to inverse 
transform the corresponding area into front view colour images of 
(80x80) pixels which we call “samples”. These samples can eas 
ily be used in Brightness and Contrast Invariant (BC-invariant) 
template matching to identify the sample in a road-sign database. 
We implemented simple BC-invariant template matching to cal 
culate a correlation score for each colour band. We calculate the 
sum of square of these three scores to obtain the general correla 
tion score. Examples are shown in figure 9. 
A A A 
/S\ A, 
^ ^ 
Figure 7: Some samples of road sign picked up from images 
thanks to fine convergence of hybrid algorithm. These samples 
can be easily used in correlation algorithm. 
AAAAAA 
AAAAAA 
AAAAAA 
AAAAAA 
AAAAAA 
AA 
Figure 8: Database of red triangular road signs used in BC- 
invariant template matching. 
6 CONCLUSION 
We present a real optimized method to fine-defect road signs in 
an image scene. Our first contribution consists in combining two 
different approaches : the first is an evolutionary algorithm which 
allows us to make global optimizations thanks to stochastic pro 
cesses, the second is a deterministic algorithm used to minimize 
the local configuration error of a deformable template which im 
proves the convergence precision and the process repeatability. 
The definition of a greater number of templates permits detection 
of other shapes such as circular road signs without no change in 
the algorithm (figure 9). Our second contribution is to combine 
the colour information and edge information to obtain an autoad- 
aptive convergence and to fine-detect objects with high precision 
with a small number of iterations. While (De La Escalera et al., 
2004) presents an analog genetic based algorithm where about 
100 individuals per template are required and more than 50 it 
erations to get their best results, our algorithm is able to con 
verge in less than 10 generations. Compared to work in (Siarry, 
2007) where different EA are used to minimize configurations of 
4 parameters, we are able to manage a 6-dimensions configura 
tion vector which enables us to extract precisely road signs with 
strong spatial tilt such as the temporary road sign in the first im 
age of figure 9. The associated sample is the first one on figure 
7 where we observe that the road sign perfectly matches the im 
age’s dimensions. Finally, we have shown that we can use simple 
Brightness and Contrast invariant template matching to recognize 
every sample that the fine-detection algorithm produces.