A Pixel Labeling Framework for Comparing Texture Features
Application to Digitized Ancient Books
Maroua Mehri
1,2
, Petra Gomez-Kr
¨
amer
1
, Pierre H
´
eroux
2
, Alain Boucher
1
and R
´
emy Mullot
1
1
L3i, University of La Rochelle, Avenue Michel Cr
´
epeau, 17042 La Rochelle, France
2
LITIS, University of Rouen, Avenue de l’Universit
´
e, 76800 Saint-Etienne-du-Rouvray, France
Keywords:
Ancient Digitized Books, Pixel Labeling, Texture, Multiresolution, Consensus Clustering, Clustering And
Classification Accuracy Metrics.
Abstract:
In this article, a complete framework for the comparative analysis of texture features is presented and evaluated
for the segmentation and characterization of ancient book pages. Firstly, the content of an entire book is
characterized by extracting the texture attributes of each page. The extraction of the texture features is based on
a multiresolution analysis. Secondly, a clustering approach is performed in order to classify automatically the
homogeneous regions of book pages. Namely, two approaches are compared based on two different statistical
categories of texture features, autocorrelation and co-occurrence, in order to segment the content of ancient
book pages and find homogeneous regions with little a priori knowledge. By computing several clustering
and classification accuracy measures, the results of the comparison show the effectiveness of the proposed
framework. Tests on different book contents (text vs. graphics, manuscript vs. printed) show that those texture
features are more suitable to distinguish textual regions from graphical ones, than to distinguish text fonts.
1 INTRODUCTION
In order to provide a wider access to libraries collec-
tions which need to be protected from too frequent
handling, numerous projects have been established.
For instance, Google has conducted large digitization
programs “Google Books Library Project” of cultural
heritage with the help of several libraries. Therefore,
automatic processing of ancient digitized documents
has undergone tremendous growth over the last few
years. Thereby, reliable ancient document interpreta-
tion systems have been the topics of major interest of
many libraries and historians and the prime issue of
research in document analysis community. There has
been a great challenge in the refinement of the well-
known approaches based on strong a priori knowledge.
Many methods have been presented in the literature
(Mao et al., 2003; Mullot, 2006) to perform this task.
However, such algorithms rely on a priori knowledge
in order to properly segment and characterize the doc-
ument image content.
The problematic of this paper concerns the segmen-
tation and characterization of ancient digitized book
content with little a priori knowledge. Specifically, in
the context of the DIGIDOC project (Document Image
diGitisation with Interactive DescriptiOn Capability)
1
,
we aim to propose new ways of interacting with scan-
ners as well as new tools for analyzing documents
during the whole acquisition process from scanning
the document to knowledge representation and man-
agement of the ancient digitized document content.
LeBourgeois et al. (Bourgeois et al., 2004) highlight
an essential necessity to conceive “intelligent” digitiz-
ers which can limit manual intervention and realize
easy and high quality digitization of image documents.
Thus, the main goal of our work is to develop a map-
ping between the scanned image and its content and
subsequently use through a set of descriptors extracted
and computed on it. Those descriptors will help repre-
senting a book page by a hierarchy of homogeneous
regions without any hypothesis on the document struc-
ture, neither on the document model nor the typograph-
ical parameters. In such conditions, a texture analysis
technique is a logical choice for solving our problem
as it gives several features characterizing the textural
properties of a region without using information on the
document structure such as the document model and
the typographical parameters (Journet et al., 2008).
1
The DIGIDOC project is referenced under
ANR-10-CORD-
0020.
553
Mehri M., Gomez-Krämer P., Héroux P., Boucher A. and Mullot R..
A Pixel Labeling Framework for Comparing Texture Features Application to Digitized Ancient Books.
DOI: 10.5220/0004804705530560
In Proceedings of the 3rd International Conference on Pattern Recognition Applications and Methods (ICPRAM-2014), pages 553-560
ISBN: 978-989-758-018-5
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
Among the most widely used texture feature ex-
traction and analysis methods are those derived from
statistical, geometrical, model-based, and signal pro-
cessing primitives (Chen et al., 1998). Several statis-
tical texture-based segmentation methods have been
presented, e.g. the autocorrelation function (Petrou and
Sevilla, 2006) and GLCM (Grey Level Co-occurrence
Matrix) (Haralick et al., 1973; Busch et al., 2005).
The extracted texture features are mainly investigated
and analyzed separately in independent experiments
for document analysis (Journet et al., 2008). Some
works deal with the whole ancient document image
(Journet et al., 2008) and others are applied to graphic
images such as drop caps (Uttama et al., 2006; Cous-
taty et al., 2011). There have been few comparative
studies on gradient, multiple channel Gabor filters, and
co-occurrence features (Payne et al., 1994; Said et al.,
2000; Liu et al., 2005; Ding et al., 2007; Zhu et al.,
2001) for document segmentation, character recogni-
tion, and script and language identification.
In (Mehri et al., 2013a), a texture-based frame-
work for segmentation of digitized historical books
is proposed. Nevertheless, the authors present only
some preliminary study of texture feature compari-
son (autocorrelation, co-occurrence, and Gabor) using
only texture feature extraction and pixel classification
tasks on simplified document images. Moreover, in
(Mehri et al., 2013c), the authors present a pixel label-
ing approach for historical digitized books based on
two non-parametric tools: the autocorrelation function
and multiresolution analysis. We propose in this paper
a complete book pixel labeling framework for com-
paring texture features based on (Mehri et al., 2013a;
Mehri et al., 2013c) including the automatic estimation
of the number of homogeneous and similar content re-
gions of book pages and the pixel labeling steps. The
framework is evaluated on a large corpus of historical
books using the autocorrelation and GLCM texture
features in order to find the homogeneous regions de-
fined by similar texture indices from the whole book
instead of processing each page individually. We com-
pare clustering and classification performance with
selectivity to the book content, i.e. text vs. graphics
and also book characteristics, such as manuscript vs.
printed, in order to summarize the pros and cons of
each texture-based method for each book category.
The remainder of this paper is structured as follows.
In Section 2, we describe the proposed framework
of pixel labeling and characterization of the content
of an entire book (Mehri et al., 2013a). The texture
primitives, autocorrelation and co-occurrence, chosen
for the validation and evaluation of the framework
are detailed in Section 3. In Section 4, we propose a
comparative analysis of the performance of the chosen
texture features. Our conclusions and future work are
presented in Section 5.
2 FRAMEWORK
A texture-based framework for segmentation of dig-
itized historical books is proposed in (Mehri et al.,
2013a). It is depicted in Figure 1. The proposed
framework is pixel-based and does not require a priori
knowledge on the document structure, neither about
the document model, nor about the typographical pa-
rameters. Thus, it is adapted to all kinds of books. It is
independent of the document layout, the typeface, the
font size, the page orientation, the digitizing resolu-
tion, etc. as a texture analysis technique is introduced.
The goal of the framework is to determine regions
or groups of pixels which share similar properties or
characteristics. These characteristics may be based on
the pixel location, their surroundings, color, intensity
or texture. In this work, we focus only on textural
characteristics. The use of an texture-based approach
has been shown to work effectively with skewed and
degraded images (Journet et al., 2008).
By selecting randomly a number of foreground
pixels from a few pages of the same book, their textu-
ral features are computed firstly. Then, an estimated
number of homogeneous and similar content regions
is computed in the analyzed book by applying the Con-
sensus Clustering method (CC) (Simpson et al., 2010)
on the extracted texture attributes (block
2
in Figure 1).
Then, for each analyzed book page its texture features
are extracted which are then used in a clustering ap-
proach by taking into consideration the estimation of
the number of clusters given before by the CC method
(block
1
in Figure 1) in order to automatically label
content pixels with the same cluster identifier with re-
spect to the book content.
Figure 1 illustrates diagrammatically the four main
tasks of our proposed framework. The block
2
on Fig-
ure 1 ensures the estimation of the number of homoge-
neous and similar content regions from the extracted
textural features on the whole analyzed book. The
block
1
on Figure 1 integrates an unsupervised task
enabling to automatically label content pixels with the
same cluster identifier regarding to the book content in
order to determine and characterize the homogeneous
regions in the digitized book (block 3 on Figure 1).
The proposed framework is described by the fol-
lowing four tasks:
1) The texture feature extraction (Section 2.1),
2) The estimation of the number of homogeneous
and similar content regions (Section 2.2),
3) The pixel clustering (Section 2.3),
ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods
554
Figure 1: Stages of our pixel labeling framework of historical
digitized book content.
4) The pixel labeling (Section 2.4).
2.1 Texture Feature Extraction
Texture feature extraction aims at representing the doc-
ument image content by a set of descriptive features
computed or extracted on it. Several studies (Uttama
et al., 2006; Journet et al., 2008) proposed to charac-
terize and index images of ancient documents by their
content by exploring textural analysis based on statis-
tical and spectral properties of texture. The authors of
(Uttama et al., 2006) introduce a segmentation method
of drop caps based on a combination of different tex-
ture analysis approaches such as the GLCM (Haralick
et al., 1973) and the autocorrelation function (Petrou
and Sevilla, 2006). In (Journet et al., 2008), the authors
propose an extraction algorithm of texture features that
is devoted to the analysis of historical documents. The
computed texture features are based on frequencies
and the autocorrelation function. This method gives
good information on the principal orientations and pe-
riodicities of the texture allowing characterizing the
content of images without assumption on the image
structure or properties.
In this paper, two statistical primitives are investi-
gated: the autocorrelation function (Petrou and Sevilla,
2006) and GLCM (Haralick et al., 1973). The extrac-
tion of textural descriptors is performed on foreground
pixels of gray-level images. The texture features are
performed at various sizes of analysis windows in or-
der to adopt a multiscale approach. A multiscale ap-
proach has been proposed by selecting concentric anal-
ysis windows with distinct sizes in order to character-
ize the images (Journet et al., 2008; Kricha and Amara,
2011). The optimal sizes of the sliding windows, re-
specting a constructive compromise between the com-
putation time and segmentation quality, have been de-
termined experimentally. The extraction of textural
descriptors is performed from only the foreground
pixels of the gray-level document images at four dif-
ferent sizes of sliding windows:
(16 × 16)
,
(32 × 32)
,
(64 × 64)
, and
(128 × 128)
. In order to avoid side ef-
fects, a border replication step is introduced allowing
computing texture features on the whole image.
2.2 Estimation of the Number of
Homogeneous and Similar Content
Regions
Since the goal of the proposed framework is to find
homogeneous regions defined by similar texture fea-
tures, a clustering approach is required in order to
partition the analyzed page into regions which have
similar properties with respect to the extracted fea-
tures. However, the number of clusters must be known
a priori especially for the conventional clustering tech-
niques (Lance and Williams, 1967; MacQueen, 1967;
Kaufman and Rousseeuw, 1990).
Previous work has identified a number of ap-
proaches (Ketchen and Shook, 1996) for determining
the correct number of clusters in a dataset. Simpson et
al. (Simpson et al., 2010) have recently proposed an
effective method, known as the Consensus Clustering
(CC), to estimate the optimal number of clusters in bi-
ological data. The idea of CC consists in performing a
consensus matrix by iterating multiple runs of cluster-
ing algorithms with random and re-sampled clustering
options (Monti et al., 2003). Therefore, the consensus
matrix analyzes the consistency of clustering results
from five different clustering algorithms: AGglomer-
ative NESting (AGNES), DIvisive ANAlysis cluster-
ing (DIANA), Partitioning Around Medoids (PAM),
k-means clustering (k-means), and Hierarchical Ascen-
dant Classification (HAC). Therefore, by weighting
the different clustering methods in order to mitigate
extremes in consensus values that can be created by
the sensitivity of some algorithms, a merge consensus
matrix is performed.
Thus, the estimation of the number of homoge-
neous and similar content regions is performed by
using the merged CC approach. Although, to perform
this task on all pages of the analyzed book is not possi-
ble in the case of huge amount of document images, be-
cause it needs a high computational time and memory.
Hence, the number of clusters is estimated by using
the merged CC method in a set of randomly selected
foreground pixels from few random selected pages of
a book. Due to the memory constraints and the high
computational time of the merged CC method, a set
of
1000
randomly selected pixels of
10
pages selected
randomly from the same book is firstly proposed.
APixelLabelingFrameworkforComparingTextureFeaturesApplicationtoDigitizedAncientBooks
555
2.3 Pixel Clustering
Since the optimal number of clusters
k
opt
is estimated,
a clustering method is required in order to characterize
the content of an entire book and find the
k
opt
ho-
mogeneous regions defined by similar texture indices
on the whole book. The work presented in (Nguyen
et al., 2010) has shown interesting results in classi-
fying the strokes of initial letters by using the HAC
algorithm. Due to the high requirement of a too large
amount of memory to perform the merged CC method
on all pixels from each analyzed document image, the
HAC algorithm is performed on the extracted textural
features without taking into account the spatial coordi-
nates and by setting the maximum number of clusters
to the estimated one,
k
opt
, with the merged CC method.
According to a hierarchical structure grouping of clus-
ters based on the criteria of the minimum increased
intra-clusters inertia, the HAC is applied on the tex-
ture features of the selected pixels of book pages. This
stage of processing gives
k
opt
clusters for the randomly
selected samples.
2.4 Pixel Labeling
This phase deals with labeling clusters or group of
pixels with respect to the results of the pixel classifica-
tion phase. The idea of this task is to assign a label to
each cluster of pixels which share similar textural char-
acteristics with respect to the obtained cluster of the
selected samples of the analyzed book. Thus, the pixel
labeling aims at determining and assigning the same
cluster identifier for each similar cluster extracted from
the digitized book.
Journet et al. (Journet et al., 2008) propose to per-
form the clustering stage by using CLustering LARge
Applications (CLARA) (Kaufman and Rousseeuw,
1990), which is known to be adapted to large scale
databases, in the extracted texture features computed
from six pages of the same book. Then, if two pixels
of two different document images have the same clus-
ter label, they belong to the same class. However, this
technique is characterized by a high computation time
and memory complexity.
In (Mehri et al., 2013b), a clustering approach with-
out taking into account the characterization step, i.e.
the pixel labeling step. In (Mehri et al., 2013a), an un-
supervised task is proposed enabling to automatically
label content pixels with the same cluster identifier
regarding to the book content. For the same book,
each cluster (represented by a given color) represents
a similar or homogeneous region. Thus, by applying
HAC, the homogeneous regions are defined by similar
texture indices. Then, the Nearest Neighbor Search
algorithm (NNS) (Knuth, 1997) is performed in order
to assign the same label for each similar cluster ex-
tracted from the digitized book. NNS is used between
each texture feature vector of each digitized page of
the same book and the
k
opt
clusters of the selected
samples of a book in order to find the closest texture
features vector to the cluster of the selected samples
of a book, i.e. by selecting the minimum distance.
Finally, the NNS (Knuth, 1997) with the Maha-
lanobis distance (Mahalanobis, 1936) is applied in
order to assign the same label for each similar clus-
ter extracted from the digitized book. The NNS is
used between each texture feature vector of each dig-
itized page of the same book and the
k
opt
clusters of
the selected samples of a book in order to find the
closest cluster to the one of the selected samples of a
book. The Mahalanobis distance takes into account
the dataset correlations and is particularly suited to
arbitrarily shaped clusters, i.e. the minimum of the
intra-cluster and inter-cluster distance is taken into
consideration.
3 COMPARISON OF
AUTOCORRELATION AND
CO-OCCURRENCE FEATURES
In order to validate and evaluate the proposed frame-
work of segmentation and characterization of ancient
digitized books, two texture primitives are computed:
the autocorrelation and co-occurrence features that are
outlined below. The different texture descriptors used
in this paper are reported in (Mehri et al., 2013a). Au-
tocorrelation and co-occurrence features are extracted
for several reasons: Firstly, we have made a compara-
tive study about choosing the texture feature category,
which ensures the best and constructive trade-off be-
tween best performance and lowest computation time.
Secondly, the pertinence of the segmentation exper-
iments (Journet et al., 2008) that are based on the
autocorrelation function applied on document image
leads us to work with the autocorrelation features in
order to reach the objectives of determining homoge-
neous regions from the analyzed document without
hypothesis on the document structure, neither on the
document model nor the typographical parameters. Fi-
nally, the extraction of the these two texture features
needs less parameter settings compared to the descrip-
tors computed from Gabor filters (Jain and Zhong,
1996) for instance. A
2
D Gabor filter is a linear se-
lective band-pass filter, dependent on two parameters:
spatial frequency and orientation. Moreover, Tamura
features depend on the value of coarseness (Howarth
ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods
556
and Ruger, 2004). Indeed, without hypothesis on the
document structure, neither on the document layout
nor the typographical parameters, the choice of ap-
propriate thresholds and parameters is a very difficult
task.
3.1 Autocorrelation Features
In this paper, we evaluate firstly the autocorrelation
descriptors. The autocorrelation function (Petrou and
Sevilla, 2006) is used to determine periodic patterns
and can characterize similarity patterns through a num-
ber of extracted autocorrelation features. A number of
autocorrelation features have been proposed in (Jour-
net et al., 2008; Ouji et al., 2011; Mehri et al., 2013b;
Mehri et al., 2013a) for segmenting ancient and con-
temporary documents images. In (Journet et al., 2008),
the authors use the directional rose (Bres, 1994), a
derivative of the autocorrelation function. The direc-
tional rose identifies significant texture orientations in
the analyzed block image.
The extracted autocorrelation descriptors provide
interesting information on the principal texture orien-
tations and periodicities. Five autocorrelation features
are computed, which have been reported in (Mehri
et al., 2013b): the main orientation of the directional
rose, the intensity of the autocorrelation function for
the main orientation, the variance of the intensities of
the directional rose, and the mean stroke width and
height estimated accurately along the axis of the main
angle of the directional rose (Journet et al., 2008; Ouji
et al., 2011; Mehri et al., 2013b).
Extracting these autocorrelation indices using a
sliding window gives a total of
20
features (
5
autocor-
relation indices
× 4
sliding window sizes for multires-
olution). Therefore, to every selected foreground pixel
from the digitized document image is assigned a vec-
tor which corresponds to the extracted autocorrelation
indices.
3.2 Co-occurrence Features
The co-occurrence attributes (Haralick et al., 1973) are
the second features tested in this paper, extracted from
the GLCM. The GLCM determines the probability of
occurrence of pixel pairs according to their gray levels
and distance by considering the spatial relationship of
pixels in the image. A GLCM element is the prob-
ability of the gray level pairs defined in a specified
direction
θ
and separated by a particular distance of
d
units. By applying multi-distance and multi-direction
approaches, a large number of co-occurrence descrip-
tors can be extracted. Fourteen textural features ex-
tracted of the GLCM have been initially introduced by
(Haralick et al., 1973) for texture discrimination of nat-
ural and satellite images. A number of co-occurrence
feature extraction and analysis methods (Mikkilineni
et al., 2005; Lin et al., 2006; Payne et al., 1994; Peake
and Tan, 1997; Busch et al., 2005) have been proposed
in order to segment and classify the content of docu-
ment images, and to identify script and language from
document images. Briefly, the GLCM matrices are ob-
tained for a small range of distance values
d = 1, 2
and
typically for the directions
θ =
{
0
◦
, 45
◦
, 90
◦
, 135
◦
}
(Busch et al., 2005).
Six co-occurrence features are extracted from the
GLCM matrices: the maximum entry in the GLCM
or the maximum probability, the correlation metric,
the energy or the angular second moment, the entropy,
the inertia or the contrast, and the local homogene-
ity for two distances
{
d = 1, 2
}
(Mikkilineni et al.,
2005; Busch et al., 2005). In addition to the twelve co-
occurrence features (six for each distance), two other
descriptors are computed: the mean value and the
standard deviation of the energy for the two distances
combined (Lin et al., 2006).
Extracting these co-occurrence indices using a slid-
ing window gives a total of
56
numeric values (
14
co-occurrence indices
× 4
sliding window sizes for
multiresolution). Therefore, to every selected fore-
ground pixel from the digitized document image a
vector is assigned which corresponds to the extracted
co-occurrence indices.
4 EVALUATION AND RESULTS
In the experiment, 316 pages extracted from 13 dif-
ferent books are considered. Our corpus is divided
into two categories: 7 printed monographs and 6
manuscripts that encompass six centuries (1200-1900)
of French history. For each category, we select three
types of page content: 110 pages containing only two
fonts, 100 pages containing graphics and single font
texts, and 106 pages containing graphics and text with
two different fonts. Evaluation of segmentation and
region classification requires a ground truth which
is performed by defining manually our ground truth
with the Ground-truthing Environment for Document
Images (GEDI)
2
, a public domain document image
annotation tool. Our corpus is composed of grayscale
pages which were digitized with 300/400 dpi. The
time required to process a page (1982*2750 pixels) us-
ing the autocorrelation approach is six minutes while
using the co-occurrence descriptors is reduced to only
one minute.
2
http://gedigroundtruth.sourceforge.net/
APixelLabelingFrameworkforComparingTextureFeaturesApplicationtoDigitizedAncientBooks
557
Figure 3 shows the real coherent separating power
of the extracted texture features in the context of histor-
ical digitized book with little a priori knowledge. For
the same book, each cluster (represented by a given
color) represents a similar or homogeneous region. Be-
cause the process is unsupervised, the color attributed
to text or graphics may differ from one book to another.
The proposed framework (cf. Figure 1) is providing
satisfying results particularly in distinguishing the tex-
tual regions from the graphical ones when comparing
visually the segmentation results for both the autocor-
relation and co-occurrence approaches. We note that
in the case of the manuscript document category (one
font and graphics), the segmentation result by the au-
tocorrelation approach (cf. Figure 3(a)) is better than
those performed by co-occurrence features (cf. Fig-
ure 3(b)), i.e. the graphic regions (blue) 3(a) are more
homogeneous. Moreover, we show for the manuscript
document category (two fonts and graphics) better re-
sults of discrimination text/graphics obtained by the
autocorrelation approach (cf. Figure 3(e)) (graphic
regions (red), textual regions (blue)) than those per-
formed by the co-occurrence approach (cf. Figure 3(f))
(graphic regions (blue), textual regions (red)). In the
case of the printed document category (one font and
graphics), the segmentation results by the autocorrela-
tion approach (cf. Figure 3(c)) show that the textural
characteristics of each small letter in the beginning of
each text line is different from the other text content
while the clustering results by the co-occurrence ap-
proach consider them as text regions (cf. Figure 3(d)).
This may be explained by the fact that the autocorre-
lation features give better information on the major
orientation and periodicities of the textual texture than
the co-occurrence ones. In general, the results for
the two approaches are relatively similar for the doc-
uments containing only two fonts when we use the
autocorrelation and co-occurrence descriptors, i.e. we
distinguish two fonts: the normal (green) and upper-
case (blue) fonts for the co-occurrence approach in
Figure 3(i) and the normal (red) and uppercase (green)
fonts for the co-occurrence approach in Figure 3(j).
We also note that in Figure 3(g), we descriminate the
normal (red) and uppercase (blue) fonts for the doc-
uments containing two fonts and graphics with the
help of the autocorrelation primitives. Nevertheless,
the co-occurrence approach offers important advan-
tages such as a reduced processing time and the ease
of implementation.
Indeed, this method of evaluating the performance
of a segmentation method is inherently a subjective
evaluation and we need to assess the effectiveness us-
ing an appropriate quantitative metric. For this reason,
we perform firstly three external, or supervised cluster-
ing, evaluation indices: Jaccard coefficient (
J
) (Saxena
and Navaneetham, 1991), Fowlkes and Mallows index
(
FM
) (Fowlkes and Mallows, 1983), and purity per
block (
PPB
) (Mehri et al., 2013b). Then, three clas-
sification accuracy metrics: precision (
P
), recall (
R
),
and classification accuracy (
CA
)(Makhoul et al., 1999)
are computed.
In Figure 2, it can be seen that the results obtained
by the numerous clustering evaluation measures are
coherent with the different classification accuracy ones
for the two approaches. We obtain
85%
,
70%
,
70%
,
and
79%
of mean purity per block accuracy, precision,
recall, and classification accuracy respectively for the
autocorrelation approach without taking into account
the topographical relationships of the selected pixels
with respect of
87%
,
70%
,
79%
, and
83%
respectively
for the co-occurrence approach.
Autocorrelation Co-occurrence
0
0,2
0,4
0,6
0,8
1
Autocorrelation Co-occurrence
0
0,2
0,4
0,6
0,8
1
Autocorrelation Co-occurrence
0
0,2
0,4
0,6
0,8
1
Autocorrelation Co-occurrence
0
0,2
0,4
0,6
0,8
1
Autocorrelation Co-occurrence
0
0,2
0,4
0,6
0,8
1
Autocorrelation Co-occurrence
0
0,2
0,4
0,6
0,8
1
C1: Manuscript-One font and graphics
C2: Printed-One font and graphics
C3: Manuscript-Two fonts and graphics
C4: Printed-Two fonts and graphics
C5: Manuscript-Only two fonts
C6: Printed-Only two fonts
C7: Manuscript
C8: Printed
C9: Overall
Jaccard (J) Fowlkes and Mallows (FM) Purity per block (PPB)
Precision (P)
Recall (R) Classification accuracy (CA)
Figure 2: Evaluation of the pixel labeling framework of
historical digitized book content by clustering and classifica-
tion accuracy measures: three clustering accuracy metrics:
Jaccard coefficient (
J
), Fowlkes and Mallows index (
FM
),
and purity per block (
PPB
), and three classification accu-
racy measures: precision (
P
), recall (
R
), and classification
accuracy (
CA
). The higher the values are, the better the
results.
We note also that the best segmentation results
are obtained by extracting the autocorrelation features
for manuscript documents containing one font and
graphics. One assumption can be that the manuscript
documents contain graphic regions that are more com-
pact and homogeneous than the printed documents
and the autocorrelation features are more suitable to
discriminate compact graphical regions from textual
ones. Whereas, the analysis of quantitative perfor-
mance noted by using the co-occurrence approach
provides different results according to the computed
accuracy. Nevertheless, the best performance is ob-
tained for documents containing graphics and text by
extracting the co-occurrence features. Concerning the
ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods
558
slight variability of the ranking of clustering perfor-
mance by using the clustering and classification accu-
racy measures can be explained by the specificity of
each clustering accuracy measure. We see that the best
classification result of the precision metric (
P
) is ob-
tained for the manuscript document category (one font
and graphics) by using both the autocorrelation and co-
occurrence approaches. Precision (
P
), recall (
R
), and
classification accuracy (
CA
) show that the lowest val-
ues are obtained for printed documents (only two fonts)
by using both the autocorrelation and co-occurrence
approaches. Therefore, the obtained qualitative results
are confirmed, i.e. the extracted autocorrelation and
co-occurrence descriptors are more suitable to distin-
guish the textual regions from the graphical ones. We
conclude that the overall results are quite satisfying
since we do not integrate the topographical or spatial
relationships.
5 CONCLUSIONS AND FURTHER
WORK
This article presents and evaluates a framework for
the comparative analysis of texture features for the
segmentation and characterization of ancient book
pages. The proposed framework is used to extract and
compare automatically texture features. Then, a non-
parametric clustering method is performed in order
to determine the homogeneous regions from different
pages of the same book. To validate the proposed
method, we evaluate two approaches based on two
different statistical categories of texture features: the
autocorrelation function and the co-occurrence matri-
ces. Our study demonstrates that the two kinds of
texture features give different results according to the
book content, i.e. text vs. graphics and also book char-
acteristics, such as manuscript vs. printed. It shows
that these texture features are more suitable to dis-
tinguish textual regions from graphical ones, than to
distinguish text fonts. However, when the numerical
complexity is taken into account, the co-occurrence
approach would be the better choice. The first aspect
of future work will be to test the framework with other
texture descriptors such as the Tamura texture features,
the local binary patterns, the multiple channel Gabor
filters, and the wavelets. Further work also needs to
be done in combining various texture descriptors in
order to construct an optimal texture feature set and
to provide a qualitative measure of which features are
most appropriate for this task.
Manuscript-One font and graphics
(a) Autocorrelation (b) Co-occurrence
Printed-One font and graphics
(c) Autocorrelation (d) Co-occurrence
Manuscript-Two fonts and graphics
(e) Autocorrelation (f) Co-occurrence
Printed-Two fonts and graphics
(g) Autocorrelation (h) Co-occurrence
Manuscript-Only two fonts
(i) Autocorrelation (j) Co-occurrence
Figure 3: Result examples of the pixel labeling framework
of historical digitized book content. For the same book, each
cluster (represented by a given color) represents a similar or
homogeneous region. Because the process is unsupervised,
the colors attributed to text or graphics may differ from one
book to another.
REFERENCES
Bourgeois, F. L., Trinh, E., Allier, B., Eglin, V., and Emp-
toz, H. (2004). Digital libraries and document image
analysis. In DIAL, pages 2–24.
Bres, S. (1994). Contributions
`
a la quantification des
crit
`
eres de transparence et d’anisotropie par une ap-
proche globale: application au contr
ˆ
ole de qualit
´
e de
mat
´
eriaux composites. PhD thesis, INSA, France.
Busch, A., Boles, W. W., and Sridharan, S. (2005). Texture
for script identification. PAMI, pages 1720–1732.
APixelLabelingFrameworkforComparingTextureFeaturesApplicationtoDigitizedAncientBooks
559
Chen, C. H., Pau, L. F., and Wang, P. (1998). Texture analysis
in the handbook of pattern recognition and computer
vision. World Scientific, second edition.
Coustaty, M., Pareti, R., Vincent, N., and Ogier, J. M. (2011).
Towards historical document indexing: extraction of
drop cap letters. IJDAR, pages 243–254.
Ding, K., Liu, Z., Jin, L., and Zhu, X. (2007). A compar-
ative study of Gabor feature and gradient feature for
handwritten chinese character recognition. In WAPR,
pages 1182–1186.
Fowlkes, E. B. and Mallows, C. L. (1983). A method for
comparing two hierarchical clusterings. JASA, pages
553–569.
Haralick, R. M., Shanmugam, K., and Dinstein, I. (1973).
Textural features for image classification. SMC, pages
610–621.
Howarth, P. and Ruger, S. (2004). Evaluation of texture
features for content-based image retrieval. IVR, pages
326–334.
Jain, A. K. and Zhong, Y. (1996). Page segmentation using
texture analysis. PR, pages 743–770.
Journet, N., Ramel, J., Mullot, R., and Eglin, V. (2008). Doc-
ument image characterization using a multiresolution
analysis of the texture: application to old documents.
IJDAR, pages 9–18.
Kaufman, L. and Rousseeuw, P. J. (1990). Finding groups in
data: an introduction to cluster analysis. John Wiley
& Sons.
Ketchen, D. J. and Shook, C. L. (1996). The application of
cluster analysis in strategic management research: an
analysis and critique. SMJ, pages 441–458.
Knuth, D. E. (1997). The art of computer programming,
volume 3: (2nd ed.) sorting and searching. Addison
Wesley Longman Publishing Co.
Kricha, A. and Amara, N. E. B. (2011). Exploring textural
analysis for historical documents characterization. JC,
pages 24–30.
Lance, G. N. and Williams, W. T. (1967). A general theory of
classificatory sorting strategies 1. Hierarchical systems.
CJ, pages 373–380.
Lin, M., Tapamo, J., and Ndovie, B. (2006). A texture-based
method for document segmentation and classification.
SACJ, pages 49–56.
Liu, C. L., Koga, M., and Fujisawa, H. (2005). Gabor feature
extraction for character recognition: comparison with
gradient feature. In ICDAR, pages 121–125.
MacQueen, J. B. (1967). Some methods for classification
and analysis of multivariate observations. In MSP,
pages 281–297.
Mahalanobis, P. (1936). On the generalised distance in
statistics. In NISI, pages 49–55.
Makhoul, J., Kubala, F., Schwartz, R., and Weischedel, R.
(1999). Performance measures for information extrac-
tion. In DARPA, pages 249–252.
Mao, S., Rosenfeld, A., and Kanungo, T. (2003). Document
structure analysis algorithms: a literature survey. In
DRR, pages 197–207.
Mehri, M., Gomez-Kr
¨
amer, P., H
´
eroux, P., Boucher, A.,
and Mullot, R. (2013a). Texture feature evaluation for
segmentation of historical document images. In HIP,
pages 102–109.
Mehri, M., Gomez-Kr
¨
amer, P., H
´
eroux, P., and Mullot, R.
(2013b). Old document image segmentation using the
autocorrelation function and multiresolution analysis.
In DRR.
Mehri, M., H
´
eroux, P., Gomez-Kr
¨
amer, P., and Mullot, R.
(2013c). A pixel labeling approach for historical digi-
tized books. In ICDAR, pages 817–821.
Mikkilineni, A. K., Chiang, P. J., Ali, G. N., Chiu, G. T. C.,
Allebach, J. P., and III, E. J. D. (2005). Printer identi-
fication based on graylevel co-occurrence features for
security and forensic applications. In SSWMC, pages
430–440.
Monti, S., Tamayo, P., Mesirov, J., and Golub, T. (2003).
Consensus Clustering: a resampling-based method for
class discovery and visualization of gene expression
microarray data. ML, pages 91–118.
Mullot, R. (2006). Les documents
´
ecrits : De la num
´
erisation
`
a l’indexation par le contenu. Herm
`
es.
Nguyen, G., Coustaty, M., and Ogier, J. M. (2010). Stroke
feature extraction for lettrine indexing. In IPTA, pages
355–360.
Ouji, A., Leydier, Y., and Bourgeois, F. L. (2011). Chromatic
/ achromatic separation in noisy document images. In
ICDAR, pages 167–171.
Payne, J. S., Stonham, T. J., and Patel, D. (1994). Document
segmentation using texture analysis. In ICPR, pages
380–382.
Peake, G. and Tan, T. (1997). Script and language identifica-
tion from document images. In DIA, pages 10–17.
Petrou, M. and Sevilla, P. G. (2006). Image processing:
dealing with texture. John Wiley & Sons.
Said, H. E. S., Tan, T. N., and Baker, K. D. (2000). Personal
identification based on handwriting. PR, pages 149–
160.
Saxena, P. C. and Navaneetham, K. (1991). The effect of
cluster size, dimensionality, and number of clusters
on recovery of true cluster structure through Chernoff-
type faces. RSS, pages 415–425.
Simpson, T., Armstrong, J., and Jarman, A. (2010). Merged
consensus clustering to assess and improve class dis-
covery with microarray data. BMC, pages 1471–1482.
Uttama, S., Loonis, P., Delalandre, M., and Ogier, J. M.
(2006). Segmentation and retrieval of ancient graphic
documents. In GREC, pages 88–98.
Zhu, Y., Tan, T., and Wang, Y. (2001). Font recognition
based on global texture analysis. PAMI, pages 1192–
1200.
ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods
560
