€ cor-
e leg-
ch the
x 100
cogni-
lage is
to en-
spares
in the
nth an
ter lo-
racter
ach of
Scores
d map
struc-
'nerate
on. To
red by
on the
ceived
1 input
ds, the
rds or-
ie lexi-
ored or
30 coin
lowest
. Top-3
itching
iscrim-
image,
se side,
e mul-
'd clas-
image.
, the set
s based
put im-
1 stride.
stering;
vocab-
ures are
Finally,
the in-
symbol.
s of the
capable
g-polar)
chieved
angular
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XL-5/W2, 2013
XXIV International CIPA Symposium, 2 — 6 September 2013, Strasbourg, France
(b)
f
Labeled sample images 9
DEM
Input: Coin Image
C Machine Learning
[ SIFT Descriptor Computation
Legend Recognition Pipeline ^
Keypoint Extraction
(384 x 384 pixels)
Resize image
Entropy filter ; Sample image
(find relevant | had ple imag
regions
image regions)
Keypoint Classification
Compute SIFT à Test against c
[mm J
(SVM
(Database
Input: Lexicon
h 4
Word Detection )
| Foe P ak M promet Le
every word orientation
N 4 y of words
[ suppor Vector Machine Learning m. sw [su mmus |
{ Database) Database
/ E T
Figure 4: (a) Training of SVMs. (b) Legend Recognition Pipeline.
Fused Coin Classification :
2
©
$ j
E 5
Obverse = 70% j
image Bg
7 60% :
Fused probability S T Fusion
Tor euch lass D 50% + Legend Recognition (Obverse+Reverse)H
sad Je Image Matching (Reverse)
4095 * - - «Image Matching (Obverse+Reverse) H
Reverse l L I i I I I I T
image
Figure 5: Overview of the fused classification method, which
combines legend recognition with global image matching.
4.4 Hybrid Coin Classification
This method fuses image matching with legend recognition in
order to improve the overall classification performance (Zam-
banini et al., 2013). This is done by computing global image
matching scores and legend word scores individually for obverse
and reverse of a coin; thus, four different scores, called pseudo-
probabilities, are computed, as illustrated in Fig. 5. Since not all
words occur on both coin sides, two different, smaller lexica can
be employed. As described before, smaller lexica lead to higher
recognition rates. Finally, the four pseudo-probabilities are com-
bined to give an overall score, which results in a better classifica-
tion rate suggesting that a combination of all methods will even
further increase the accuracy. Fig. 6 shows the accuracy for im-
age matching, legend recognition and the hybrid method. The
x-axis shows the number of classes N within which the correct
class was recognized; i.e., for N — 5, the correct class is found
among the 5 classes with the highest computed overall probabil-
Ity.
5 CONCLUSIONS AND FUTURE WORK
This paper presented the ILAC project, in the course of which
several image-based classification methods for ancient Roman
Republican coins have been developed. Different aspects of the
depicted imagery, such as the legend or the copin image, are
recognized individually for an efficient preselection of classes.
This way, computationally more complex classification methods
considering the entire coin only need to be performed for a sub-
set of sample classes. A first approach towards fusing the dif-
ferent classification methods has already been implemented and
377
5.10: 15 20 25 30 35 40 45 50. 55. 60
N
Figure 6: Classification accuracy of legend recognition, image
matching and hybrid coin classification for the top N classes.
showed promising results, which encourages a further combina-
tion of even more individual recognition techniques. Besides the
improvement of the classification accuracy of the existing meth-
ods, the reduction of processing time will be tackled within the
next few months. The ultimate goal is to provide a framework
that generates an ordered list of classes sorted by how well they
match the given input image that can be used in various applica-
tions and hides the underlying complexity from the user. Appli-
cation areas range from an online web portal for the numismatic
community to applications for mobile devices which allow for an
image-based coin classification with smartphones.
ACKNOWLEDGEMENTS
This research is funded by the Austrian Science Fund (FWF):
TRP140-N23-2010.
REFERENCES
Anwar, H., Zambanini, S. and Kampel, M., 2013. Supporting
ancient coin classiffcation by image-based reverse side symbol
recognition. In: CAIP 2013 - The 15th Conference on Computer
Analysis of Images and Patterns.
Arandjelovic, O. D., 2010. Automatic attribution of ancient ro-
man imperial coins. In: Proceedings of the Conference on Com-
puter Vision and Pattern Recognition, pp. 1728-1734.
Arandjelovic, O. D., 2012. Reading ancient coins: automatically
identifying denarii using obverse legend seeded retrieval. In: Pro-
ceedings of the 12th European conference on Computer Vision -
Volume Part IV, pp. 317—330.
