ROBUST AUTOMATIC SEGMENTATION OF ANCIENT COINS

Sebastian Zambanini and Martin Kampel

Pattern Recognition and Image Processing Group, Vienna University of Technology, Favoritenstr. 9/1832, Vienna, Austria

Keywords:

Segmentation, Shape Description.

Abstract:

Nowadays, ancient coins are becoming subject to a very large illicit trade. Thus, the interest in reliable

automatic coin recognition systems within cultural heritage and law enforcement institutions rises rapidly.

Central component in the permanent identiﬁcation and traceability of coins is the underlying image recognition

technology. Prior to any analysis a coin image has to be segmented into two areas: the area depicting the coin

and the area belonging to the background. In this paper, we focus on the segmentation task as a preprocessing

step for any automated coin recognition system. The objective is a robust segmentation procedure for a large

variety of coin image styles. We present a simple and fast method for coin segmentation, based on local

entropy and gray value range. Results of the developed algorithm are shown for an image database of ancient

coins and demonstrate the beneﬁts of our approach.

1 INTRODUCTION

Traditional methods to ﬁght the illicit trafﬁc of an-

cient coins comprise manual periodical search in auc-

tions catalogues, ﬁeld search by authority forces, pe-

riodical controls at specialist dealers, and a cumber-

some and unrewarding internet search, followed by

human investigation. Therefore, image-based meth-

ods to automatically recognize ancient coins have the

potential to increase their traceability to a high degree

and thus to help to combat their illicit trade.

For the image-based recognition of ancient coins,

initially a segmentation of the coin region is of out-

most importance. Especially for the identiﬁcation of

stolen coins a correct segmentation is a crucial step

since the shape of the coin provides a substantial fea-

ture to identify a concrete coin specimen (Zaharieva

et al., 2007). An automatic segmentation method is

also of great beneﬁt for the indexing of new coins

since up to now numismatists have to perform this

time-consuming task manually.

In the context described above, the methods have

to deal with images from various sources (e.g. mu-

seum collections or public online databases). There-

fore, no assumptions about image quality can be made

and major challenges that have to be faced in the seg-

mentation of coins are caused by an improper im-

age acquisition procedure. Especially shadow casts

caused by an insufﬁcient illumination setup impede

the correct determination of the coin border. Further-

more, tests have shown that image compression with

chroma subsampling is often used when storing im-

ages of coins. The resulting compression artifacts

preclude the use of color information, thus only the

luminance can be used for a reliable segmentation of

the coins.

In this paper a simple and fast method for coin

segmentation based on local entropy and gray value

range is presented. The remainder of this paper is or-

ganized as follows. Related work and the coin seg-

mentation strategy itself are addressed in Section 2.

Experiments on a set of 92 images are reported in Sec-

tion 3. A conclusion is ﬁnally given in Section 4.

2 COIN SEGMENTATION

Coin segmentation deals with the division of the im-

age into two regions: the region depicting the coin

and the region belonging to the background. In (Za-

harieva et al., 2007) segmentation of ancient coins

was achieved using an adaptive thresholding method

originally suggested by (Yanowitz and Bruckstein,

1989). The proposed method derives a threshold sur-

face obtained by an interpolation of tie points placed

at thresholded gradient values. However, as demon-

strated in the experiments (Section 3), this method

fails if the coin images show a high variability.

Segmentation of present day coins was done in

273

Zambanini S. and Kampel M. (2009).

ROBUST AUTOMATIC SEGMENTATION OF ANCIENT COINS.

In Proceedings of the Fourth International Conference on Computer Vision Theory and Applications, pages 273-276

DOI: 10.5220/0001798302730276

 SciTePress

various papers. However, all of them make special as-

sumptions which are not satisﬁed in our image data:

(Reisert et al., 2006) apply the Hough transformation

for circle detection. By deﬁnition, this approach is not

applicable on ancient coins which likely show no per-

fect circularity. The global thresholding methods pre-

sented in (van der Maaten and Poon, 2006) and (N

olle

et al., 2003) are applied to images acquired under con-

trolled conditions and are therefore not appropriate to

segment images from many different sources.

Because of the problems stated, we propose a ro-

bust method which is able to correctly segment a vari-

able set of different coin images. The only assump-

tion we make is that the coin itself possesses more

local information content and details than the rest of

the image, i.e. the background. For that reason, our

method is based on two ﬁlters providing a local mea-

surement of information content in the image: the lo-

cal entropy and the local range of gray values.

Local Entropy. Entropy is the measure of the in-

formation content in a probability distribution. For

digital images the probability distribution is repre-

sented by the histogram of gray values (Kapur et al.,

1985). If Ω deﬁnes a local neighborhood within the

image with gray value frequencies p

, p

, ..., p

(i.e.

the normalized histogram values), the local entropy is

deﬁned as

H(Ω) = −

∑

k=1

· log

) (1)

Local Range of Gray Values. The local range of

gray values is deﬁned as the difference of the max-

imum and minimum gray value of a local neighbor-

hood.

The outputs of these two ﬁlters are summed-up to

build the ﬁnal intensity image where the thresholding

is applied on. For both ﬁlters a circular neighborhood

with a radius of 3 pixels is used and both ﬁlter outputs

as well as the ﬁnal intensity image are normalized to

the range 0 to 1. For illustration on a simple example,

in Figure 1 the particular results of the entropy ﬁlter

(b), the range ﬁlter (c) and their summation (d), ap-

plied to a coin image (a), are shown. The output of

both ﬁlters is higher for the region of the coin than for

the region of the background, especially at the coin

border.

To obtain the ﬁnal coin segmentation from the in-

tensity image shown in Figure 1d, a simple way would

be to apply a global threshold and close all holes in

the binary mask caused by homogeneous regions in-

side the coin. However, tests have shown that such a

manually deﬁned threshold does not perform well on

(a) (b)

Figure 1: (a) original image, (b) output of local entropy ﬁl-

ter, (c) output of local range ﬁlter, (d) sum of local entropy

and local range (ﬁnal intensity image).

the overall given test set. Therefore, a more sophisti-

cated approach is used: we apply seven thresholds T

= 0.3, 0.35, ..., 0.6) to the normalized intensity im-

age and compute a score for each achieved segmen-

tation that represents the conﬁdence to the given seg-

mentation. Afterwards the segmentation with highest

conﬁdence is chosen.

(a) Original image. (b) T

= 0.3, formfactor

= 0.280

= 0.4, formfactor =

0.708

(d) T

= 0.5, formfactor

= 0.812

Figure 2: Four segmentation masks according to different

thresholds T

applied to the intensity image.

Since the shape of a coin is close to a circle, we

use the formfactor (Russ, 2006) of the binary segmen-

tation mask as conﬁdence measure. The formfactor of

a binary mask is computed as follows:

formfactor =

4πA

(2)

where A is the area and P the perimeter of the binary

VISAPP 2009 - International Conference on Computer Vision Theory and Applications

274

mask. The formfactor is sensitive to both the elonga-

tion of a region and the jaggedness of its border. The

higher the jaggedness of the border, the less the form-

factor. The formfactor is equal to 1 for a circle and is

less for any other shape. Since the ﬁnal shape of the

segmentation should be close to circle with a regular

border, the formfactor provides a convenient measure

for the conﬁdence of the segmentation.

Since low thresholds can produce a coin segmen-

tation that is near the rectangular shape of the whole

image (providing a comparatively high formfactor), a

segmentation is furthermore only accepted if the area

of the segmented region is lower than 90 % of the

image area. In the case that thresholding produces

more than one connected component in the image,

the one with highest formfactor covering at least 5 %

of the image area is selected. An example for differ-

ent segmentations obtained with different thresholds

is shown in Figure 2. The segmentation obtained with

= 0.5 shown in Figure 2d has the highest formfac-

tor and is therefore chosen as the ﬁnal segmentation.

The discretization of T

= 0.3, 0.35, ..., 0.6 was

chosen empirically. Tests have shown that a ﬁner

discretization does not improve the accuracy of the

method.

3 EXPERIMENTS

The proposed method was tested on a set of 92 images

acquired at the Kunsthistorisches Museum Vienna,

Austria, the Fitzwilliam Museum, Cambridge UK, and

the Romanian National History Museum representing

a wide range of different coin images. The images

differ in various ways (resolution, background, coin

size relative to image size, illumination conditions).

For the experiments presented here, all color im-

ages were converted to gray-level images. Compres-

sion artifacts due to chroma subsampling are highly

present in the data and make the use of color informa-

tion infeasible.

For each image a ground truth segmentation was

manually obtained by means of a commercial image

editing program. For the evaluation of the segmen-

tation the mutual overlap (MO) (Bowyer, 2000), also

known as dice coefﬁcient, is measured:

MO =

2 · |A

∩ A

| + |A

(3)

where A

is the set of pixels in the segmented re-

gion and A

the set of pixels in the ground truth seg-

mentation.

To demonstrate the appropriateness of the pro-

posed method for the segmentation of coin images,

the results are compared to the outputs of various

other segmentation methods: (1) the adaptive thresh-

olding method used in (Zaharieva et al., 2007) for

the segmentation of ancient coins, (2) the mean shift

method proposed by (Comaniciu and Meer, 2002) for

a comparison with a state-of-the-art method in image

segmentation and (3) our method when the threshold-

ing is directly applied to gray values instead of the

sum of entropy and gray value range.

It must be noted that the output of the mean

shift segmentation method is not implicitly a parti-

tion into foreground and background, as needed here.

Mean shift partitions the image in a set of disjoint

regions without labeling the foreground and back-

ground. From our point of view the segmentation has

to extract the single most salient object in the image,

which is in our case the coin. Therefore, to make the

mean shift segmentation results comparable, the pa-

rameter M for the minimum allowable region area has

to be manually adapted for each image to produce a

two-segment partition of the image. Evaluation was

performed on the mean shift implementation of the

EDISON system

In Table 1 the average and median MO for the dif-

ferent methods are listed. The average MO of 0.517

and median MO of 0.720 of the adaptive threshold-

ing method indicate its low robustness. Although the

parameters of the method can be adjusted to perform

well on a given type of coin image it is not able to

handle the wide range of different images contained

in the test set. A second conclusion of the results

is that the local entropy and range ﬁltering is a rea-

sonable preprocessing step to provide a more appro-

priate intensity image for the thresholding. This can

be seen by the lower average and median MO when

the original gray values are used. From the results in

Table 1 it can also be seen that our method achieves

a similar performance than the state-of-the-art mean

shift segmentation. The average MO is equal (0.983)

and the median MO of our method is even higher

(0.993 to 0.988 of the mean shift method). However,

our method has two advantages: ﬁrstly, in contrast to

mean shift no parameter has to be adapted manually.

And secondly, our method is computationally faster:

our method (written in MATLAB 7.1) takes 0.38s for

a 178 × 184 image and 8.40s for a 1154 × 866 image,

whereas the means shift implementation (written in

C++) takes 0.73s for the 178 × 184 image and 29.37s

for the 1154 × 866 image on the same machine.

Figure 3 shows results on selected images where

the obtained coin border is outlined by a black or

white line. Figure 3a-c belong to the best segmenta-

http://www.caip.rutgers.edu/riul/research/code/EDISON/

index.html, last visited: November 18th 08

ROBUST AUTOMATIC SEGMENTATION OF ANCIENT COINS

275

Table 1: Average and median MO achieved on the 92 test

images.

Average Median

Adaptive Thresholding 0.517 0.720

Mean Shift 0.983 0.988

Our method on original gray values 0.923 0.980

Our method 0.983 0.993

(a) MO = 0.9973 (b) MO = 0.9981

(e) MO = 0.9500 (f) MO = 0.9904

Figure 3: Results of the proposed segmentation method.

tion results with a MO of 0.9973, 0.9981 and 0.9970,

respectively.

Figure 3d-e belong to the worst results with a MO

of 0.9441 and 0.9500, respectively. You see that shad-

ows pose a problem to the method since they produce

a strong edge not belonging to the actual coin bor-

der. However, on the image of Figure 3f the method

correctly excludes the shadow from the segmentation,

producing a MO of 0.9904.

4 CONCLUSIONS

The method shows convincing results with a median

MO of 0.9928 and proves that local entropy and gray

value range give a convenient estimate of the actual

coin region. However, although the method’s robust-

ness is indicated by a minimum MO of 0.9048 on a set

of 92 test images, shadows still pose a problem. Nev-

ertheless, the method outperforms the state-of-the-art

mean shift segmentation method both in segmenta-

tion accuracy and speed. Furthermore, our method

needs no parameter adjustment and is therefore able

to deal with a large variety of coin image styles. To

sum up, the results achieved satisfy the needs of au-

tomatic coin identiﬁcation for a large variety of coin

image styles, nevertheless future research focuses on

the segmentation accuracy in the occurrence of shad-

ows.

ACKNOWLEDGEMENTS

This work was partly supported by the European

Union under grant FP6-SSP5-044450. However, this

paper reﬂects only the authors’ views and the Euro-

pean Community is not liable for any use that may be

made of the information contained herein.

REFERENCES

Bowyer, K. B. (2000). Validation of medical image anal-

ysis techniques. In Handbook of Medical Imaging,

volume 2, pages 567–607. Press Monograph.

Comaniciu, D. and Meer, P. (2002). Mean shift: A robust

approach toward feature space analysis. IEEE Trans.

Pattern Anal. Mach. Intell., 24(5):603–619.

Kapur, J. N., Sahoo, P. K., and Wong, A. K. C. (1985). A

new method for gray-level picture thresholding using

the entropy of the histogram. CVGIP, 29:273–285.

olle, M., Penz, H., Rubik, M., Mayer, K. J., Holl

ander,

I., and Granec, R. (2003). Dagobert – a new coin

recognition and sorting system. In Proc. of DICTA’03,

pages 329–338.

Reisert, M., Ronneberger, O., and Burkhardt, H. (2006). An

efﬁcient gradient based registration technique for coin

recognition. In Proc. of the Muscle CIS Coin Compe-

tition, pages 19–31.

Russ, J. C. (2006). The Image Processing Handbook. CRC

Press, 5th edition.

van der Maaten, L. J. and Poon, P. (2006). Coin-o-matic: A

fast system for reliable coin classiﬁcation. In Proc. of

the Muscle CIS Coin Competition, pages 07–18.

Yanowitz, S. and Bruckstein, A. (1989). A new method for

image segmentation. CVGIP, 46(1):82–95.

Zaharieva, M., Huber-M

ork, R., N

olle, M., and Kampel, M.

(2007). On ancient coin classiﬁcation. In Proc. of

VAST’07, pages 55–62.

VISAPP 2009 - International Conference on Computer Vision Theory and Applications

276