In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009 
Figure 9: The system localizes correctly text in the image 
(even with rotated text) but it detects aligned windows as 
text. 
Figure 10: Text is correctly localized, but the classification 
step fails on the end of the word courcmt in red and zebra 
crossing sign is seen as text. 
We also test different combinations of classifiers and de 
scriptors. When we try early fusion architecture, we give 
all descriptors to a unique svm classifier ; the result does 
not even reach 74% of accuracy. On the contrary, if we 
add a collection of simple geometric descriptors (compac- 
ity, surface, concavity...) to the svm classifier that must 
take the final decision in our architecture, the overall ac 
curacy reaches 88,83%. These measures seem to help the 
classifier to select which classifiers are the most reliable 
depending on the situation. 
The overall accuracy seems to be a bit low but the vari 
ability of text in our context is so huge that the real perfor 
mance of the system is not so bad. 
8 TEXT LOCALIZATION IN CITY SCENES 
Let us see the application of the complete scheme. We took 
an initial image (Figure 12). The application of our algo 
rithm of segmentation gives the result in figure 13. All re 
gions with a reasonable size are kept, others are dismissed 
(Figure 14). The classifier selects text regions among re 
maining regions (Figure 15). Text regions are grouped to 
create words and sentences (Figure 16). 
The system is efficient: instead of a variation of orienta 
tion, police and lighting condition, the system handles ma 
jority of text (Figure 9, 10 et 11). But it also generates 
many false positives: especially aligned windows (Figure 9 
top right and Figure 11). Other results can be seen in fig 
ures 9 and 10. The system must then be improved to reduce 
false positives. 
ALIMENTATION GENERALE 
»0143720884 
Figure 11: Various texts are correctly handled but periodi 
cal features are also interpreted as text. 
9 CONCLUSION 
We have presented a text localization process defined to 
be efficient in the difficult context of the urban environ 
ment. We use a combination of an efficient segmentation 
process based on morphological operator and a configu 
ration of svm classifiers with various descriptors to deter 
mine regions that are text or not. The system is competi 
tive but generates many false positives. We are currently 
working to enhance this system (and reducing false posi 
tives) by improving the last two steps: we keep on testing 
various configurations of classifiers (and selecting kernels 
of svm classifiers) to increase the accuracy of the classi 
fier and we are especially working on a variable selection 
algorithm. We are also working on the grouping step of 
neighbour text regions and its correction to send properly 
extracted text to O.C.R. 
ACKNOWLEDGEMENTS 
We are grateful for support from the French Research Na 
tional Agency (A.N.R.) 
REFERENCES 
Arth, C., Limberger, F. and Bischof, H., 2007. Real-time license 
plate recognition on an embedded DSP-platform. IEEE Interna 
tional Conference on Computer Vision and Pattern Recognition 
(CVPR ’07) pp. 1-8. 
203