Decorated initial segmentation
Je´rome Landre´, Fre´de´ric Morain-Nicolier, Lafi Alotaibi, Su Ruan
Univ. de Reims-Champagne-Ardenne - CReSTIC
Institut Universitaire de Technologie
9 rue de Que´bec, 10026 Troyes Cedex, France,
frederic.nicolier@univ-reims.fr
Abstract. This article deals with document images analysis. Nowadays,
large databases of ornaments of the hand-press period are available and
need efficient retrieval tools for history specialists and general users. In
both cases, automatic retrieval systems will help them to access the or-
nament databases. We focused in this paper on decorated initials. The
end-objective is the recognition of several metadata associated to the ini-
tial letters. This paper deals with the segmentation of the initial letter
represented in the image. The decorations consist mainly in small size in-
formation. The proposed algorithm consists in 1/ a multiresolution anal-
ysis providing a low resolution version of the image, 2/ a filtering that
removes small particles (hypothezised to be decorations) and finally 3/ a
coarse to fine progressive improvement of the segmentation by successive
applications of step 2. The algorithm is validated by its application on a
small set of representative images.
Keywords. segmentation, multiresolution analysis, ornamental letters,
mathematical morphology, coarse-to-fine.

1 Introduction
During the last 25 years many works have been done on document images re-
trieval concerning official forms, maps, drawings, correspondences, etc. We focus
here on the segmentation of ornaments from the Hand-Press period. This period
runs from around 1454 (approximate date of Gutenberg’s invention) to through
the first half of the nineteenth century (when mechanized presses started to ap-
pear). The particularity of this period is the use of blocks of wood, with a relief
carving on it, to print the ornaments of the books. Some examples of such orna-
ments are presented in figure 1. The proposed works take part of the Calypod
research group interests. Calypod is a research group interested with the image
retrieval applied to printed old books. Especially, the main topics of Calypod
concern processing of graphical parts inside these books like headpieces, pictures
or ornamental letters (also called initial letters in this article). It investigates the
use of computer science techniques in order to pre-process, to segment, to rec-
ognize, to retrieve or to match these graphical parts from these books [3].
Fig. 1. (from left-right and top-down) decorated initial, printing-house trademark, her-
aldry, emblem, picture, fleuron ornaments. We focused in this paper on the decorated
initials.
With the growing of interest in cultural heritage preservation in the 1990s,
numerous works of digitization of historical collections have been carried out.
Nowadays, large databases of ornaments of the hand-press period are available
(see Tab. 1) and will still grow in the future. These ornaments are extracted
from the whole digitized pages, using full automatic [8] or user-driven [12] seg-
mentation methods, or recorded independently. Previous works of the Calypod
research group have emphasized that the key problem is now to make available
these images for research to history specialists and for general users (like artists,
designers, publishers, printers, students, etc.). All of these constituencies have
different needs which requires varied and sophisticated means of searching and
accessing information.
General users expect web retrieval systems. They want to retrieve ornaments
using text queries but also content-based ones to access items, textures, colors,

Name size main type period (century)
HERON [4] 32 000 heraldry 18
BVH [1] 10 000 initials, emblems 16
Passe-partout [10] 4 000 fleurons, headlines 16-18
Table 1. Some databases of ornaments.
etc. Example of such an information system is the Hermitage Museum: it allows
text and content-based queries to retrieve pictures according to metadata, color
distribution and layouts. History specialists have different needs. They want
to record the individual instances of ornament occurrence in order to identify
individual blocks. Once blocks are identified, they constitute thesaurus using
subject-specific classification system (like Iconclass [5]) to describe blocks. In
both cases, automatic retrieval systems will help users to access the ornament
databases. The development of such systems is a challenging task for the Docu-
ment Image Analysis community due to different key problems:
– image degradations (old age, lossless compression)
– scaling invariance (different scanning resolutions)
– complexity (masses of data)
– scalability (high number of block classes)
In the last 10 years several works have been proposed on this topic. An
overview can be found in [PAR 06a] but dedicated to a specific collection of
methods (i.e. developed during a specific project and applied to a single corpus).
We focused in this paper on decorated initials (see figure 1 and figure 2). The
end-objective is the recognition of several metadata associated to the initial
such as : letter (”a”, ”b”, ”c”, ...), color (black, white), alphabet (latin, greek,
hebraic), style (roman, gothic).
Fig. 2. Some decorated initials (also called ornamental letters).
This paper deals with the segmentation of the initial in the image. This is
a hard task due to the high level of decoration of the initials. Images are also
highly-constrasted, resulting in no trivial difference between the initial letter
itself and background decorations. Decorations are mainly small size information
compared to the initial by itself. The first step of the segmentation process is
thus a multiresolution analysis of the image in order to have a nice represention

of the small and big spatial information. At high scales only big information is
retained and we make the hypothesis that the initial spatial information is mainly
coarse. The second step of the segmentation process is the selection of a working
scale, providing an approximation of the image. This approximation needs to
be binarized and filtered in order to keep only the biggest region of connected
pixels (connected component). The rest of the paper is organized as follows: our
segmentation process is presented in section 2 and results are detailed in section
3. We conclude with some remarks about the future work needed to improve the
proposed segmentation.
2 Segmentation of the initials
In order to obtain a good segmentation, the background information have to be
removed from the ornamental letter image. The idea of this work is to use mul-
tiresolution analysis as a tool to remove small details of the image to keep only
the shape of the letter for processing. The principle of our method is proposed
in figure 3. It starts with the original image and leads to the segmentation result
which shows a better representation of the letter.
Fig. 3. Principle of our system.
The original ornamental letter image (a) is transformed to a multiresolution
image (b). The approximation image (c) is extracted from image (b). After
binarization, a binary image representing the letter is obtained (d). The biggest
region of connected pixels is selected from (d) laeding to image (e). At last, a
binary reconstruction of seed (a) and mask (e) gives the result image (f) which
is supposed to be the segmentation image of the letter.

2.1 Multiresolution Analysis
Multiresolution analysis is a powerful tool for analysing images at multiple scales.
It has been used in many fields of image processing for two decades: compression,
denoising, coding, analysis, features extraction, indexation. . . In our work, mul-
tiresolution analysis was chosen because of its ability to separate details from
approximation leading to a multiresolution representation of images. Further-
more it allows a coarse to fine progression in order to incrementally improve the
quality of the segmentation.
An implementation of the integer lifting scheme algorithm published by
Calderbank et al. [2] have been developped. In this method, only integer num-
bers are used during transformation computing. Working with integers ensures
a very efficient computing time and a very simple algorithm.
Figure 4 shows an example of a two-level multiresolution analysis. In the
obtained low level approximation image, the shape of the represented letter
appears better than in the original image because a lot of details of the image
background have been removed from approximation to details.
(a) (b) (c)
Fig. 4. Multiresolution analysis example (a) original image (b) multiresolution decom-
position at scale 2 and (c) approximation (zoomed) at scale 2.
2.2 Binarization
The aim of binarization is to keep most of the letter pixels while removing most
of the decoration pixels. Binarization is realized thanks to the Otsu algorithm
[9]. Under the hypothesis that the image histogram can be expressed as the sum
of two gaussians, the histogram is divided in two classes. The inter-class variance
is then minimized. An example of the Otsu thresholding is given in figure 5.
As shown in figure 5, the raw Otsu binarization keeps to much pixels in the
decoration area. We thus modify the algorithm in order to offset the threshold.
The offset T is a relative value computed as follows :
T = α(M − t) + t (1)

Fig. 5. An example of Otsu thresholding an decorated initial (on the left). On the right
the binarized image contains to much pixels in the decoration area.
where m and M are respectively the minima and the maxima of the pixel
gray-level values in the image, t is the threshold given by the Otsu algorithm.
α is the relative offset, α ∈ [0, 1]. If α = 0 then T = t, the threshold equals the
Otsu threshold. If α = 1 then T = M , the threshold equals the maxima of the
image (i.e no pixels remains after thresholding).
The numerical value of α will be empirically evaluated in section 3 with the
objective to have the same offset α for all the processed images. An example of
the obtained binarization is given in figure 6.
Fig. 6. An example of Otsu thresolding with an offset α = 0.6. The number of pixels
in the decoration area is slightly reduced. The letter is quite deteriorated, but will be
improved by the filtering step (see section 2.3).

2.3 Filtering small components
Here, the objective is to remove decorations while the initial is mostly preserved.
In mathematical morphology this is known as blob analysis. A way to perform a
blob analysis is a morphological reconstruction, more specifically a reconstruc-
tion by dilation.
Reconstruction by dilation allows reconstruction of all objects of an image
X marked by an image Y . This a very powerful morphological operation that
reconstruct (retains) connected particles in an image (called mask) based on
markers present in another image (called seed). Morphological reconstruction
consists in dilating the seeds inside the mask (so particles that do not have seeds
are not reconstructed). The propagation of the seeds is restricted to the particles
they belong to [13, chap. 6 - Geodesic Transformations].
It is thus needed to have a mask image and a seed image. The mask image will
be the binarized image obtained from 2.2. The seed image needs to be idealy an
image containing only pixels from the letter. As we make the hypothesis that the
initial spatial information is mainly coarse, the decoration pixels can be removed
from the binarized image thanks to an erosion operator.
The final filtering algorithm is an erosion, followed by a binary reconstruction
of the binarized image based on the eroded image as the seed. An application of
this filtering is given in figure 7.
(a) (b) (c) (d)
Fig. 7. An example of removing small particles. (a) is a decorated initial. (b) is the
image binarized. (c) is obtained from an erosion with a squared structured element of
radius 3. (d) is the binary reconstruction with mask (b) and seed (c).
3 Results
It will be shown further that the proposed algorithm needs some minor improve-
ments. Thus a small representative subset of the whole image database is used
here to highlight the main behaviour of the segmentation. Six images are used
for the test and are shown in figure 8. These six images show various difficulties.
An implementation of our method has been programmed under the ImageJ
[6] Java image processing platform. The developped plugins can be downloaded

Fig. 8. Test images.
from the authors’ website [11] Integer lifting scheme algorithm and Otsu en-
hanced binarisation method were developped from scratch while erosion and
reconstruction were adapted from existing ImageJ plugins [7].
The parameter α (eq. 1) which corresponds to the threshold offset in the
Otsu binarisation method was fixed using the experimental binarisation result
on several images of the database. The values 0.4, 0.5, 0.6 and 0.7 were tried
for α and the best results were found for α = 0.6. During the erosion and
reconstruction steps, a structuring element with a ”plus” shape of three pixels
height and three pixels width have been used. The 4-connexity model have been
used for mathematical morphology operators. These choices about 4-connexity
and structuring element size have been proved to give the best results on the
ornamental letters database.
Fig. 9. Segmentation results on test images.
Figure 9 shows the segmentation results for the six chosen images. Results are
very good, even on difficult images. Two initials are perfectly segmented : ”A”
and ”M” is spite of the bad quality of the ”M” (the image contains information
of the verso bookpage). The image of initial ”G” contains figurative decorations
of big size and thus present a challenge to the segmentation process. The result
is an almost well segmented initial as only a small connected blob remains on the
left of the letter. The two ”C”s are well segmented if we ignore the border. For
the last letter ”V”, several small decoration parts still remain. This is mainly
due to the bad quality of the image as the printed letter is quite damaged.
4 Conclusion
In this article, an original method segmentation method for ornamental let-
ters has been proposed. This method consists in a coarse to fine segmentation
with progressive improvements. The atomic step of the process is composed of

a modified Otsu binarization, followed by a selection of the biggest connected
component. This image represents the seed of a morphological reconstruction.
Our multiresolution approach offers a very good separation of letter informa-
tion and decoration information. Binary morphological reconstruction gives very
good result images. Results are very good even on difficult images composed of
many decorations of various sizes which are the most difficult images to process
in such kind of image analysis problems.
Future works will consist in small post-processing in order to clean result
images: remove remaining borders (if needed) and small components. The next
step will be then to perform the letter recognition. The application of the whole
process to a large set of image is also planned.
References
1. Bibliothe`ques Virtuelles Humanistes: http://www.bvh.univ-tours.fr
2. Calderbank, A. R., Daubechies, I., Sweldens, W., Yeo, B.: Wavelet transforms that
map integers to integers Journal of Wavelet transforms that map integers to integers
5 (1998) 332–369
3. Calypod research group: http://calypod.free.fr
4. HERON image retrieval system: http://www.informatik.uni-augsburg.de/heron
5. IconClass classification system: http://www.iconclass.nl
6. Rasband, W.S.: ImageJ, U. S. National Institutes of Health, Bethesda, Maryland,
USA http://rsb.info.nih.gov/ij (1997-2008)
7. Landini, G.: Morphology collection of ImageJ Java plugins
http://www.dentistry.bham.ac.uk/landinig/software/software.html (2008)
8. Journet, N., Ramel, J.-Y., Mullot, R., Eglin, V.: A proposition of retrieval tools
for historical document images librairies International Conference on Document
Analysis and Recognition (ICDAR) 2 (2007) 1053–1057
9. Otsu, N.: A threshold selection method from gray level histograms IEEE Trans.
Systems, Man and Cybernetics 9 (1979) 62–66
10. Passe-partout image retrieval system: http://www2.unil.ch/BCUTodai/app
11. Nicolier, F.: Image processing resources http://pixel-shaker.fr
12. Ramel, J.-Y., Leriche, S., Demonet, M.-L., Busson, S.: User-driven page layout
analysis of historical printed books International Journal on Document Analysis
and Recognition (IJDAR) 9 2–4 (2007) 243–267
13. Soille, P.: Morphological image analysis: principles and applications (2003) Springer