Handwritten Arabic text line segmentation using affinity ... - CiteSeerX

Handwritten Arabic Text Line Segmentation using Affinity ∗ Propagation Jayant Kumar

Wael Abd-Almageed

Le Kang

David Doermann

Language and Media Processing Laboratory Institute of Advanced Computer Studies University of Maryland College Park, MD, USA

{jayant,wamageed,lekang,doermann}@umiacs.umd.edu ABSTRACT

1. INTRODUCTION

In this paper, we present a novel graph-based method for extracting handwritten text lines in monochromatic Arabic document images. Our approach consists of two steps Coarse text line estimation using primary components which define the line and assignment of diacritic components which are more difficult to associate with a given line. We first estimate local orientation at each primary component to build a sparse similarity graph. We then, use a shortest path algorithm to compute similarities between non-neighboring components. From this graph, we obtain coarse text lines using two estimates obtained from Affinity propagation and Breadth-first search. In the second step, we assign secondary components to each text line. The proposed method is very fast and robust to non-uniform skew and character size variations, normally present in handwritten text lines. We evaluate our method using a pixel-matching criteria, and report 96% accuracy on a dataset of 125 Arabic document images. We also present a proximity analysis on datasets generated by artificially decreasing the spacings between text lines to demonstrate the robustness of our approach.

Text line segmentation is a critical preprocessing step in document analysis [1] and is especially difficult for handwritten material. Historically, text lines are crucial for analyzing the document layout [2], assessing the skew or orientation of a document [3, 4] and indexing/retrieval based on word and character recognition [5]. Although text line segmentation for machine printed documents is often seen as a solved problem, freestyle handwritten text lines still present a significant challenge [9, 17]. This is because handwritten text lines are often curved, have nonuniform space between lines and have spatial envelopes that may overlap. Irregular layout, variable character sizes originating from different writing styles, the existence of touching lines and the lack of a well defined baseline also contribute to making handwritten document analysis more difficult [11]. For Arabic, the presence of diacritical components further complicates the task. Existing line extraction techniques may be categorized as global projection based, local grouping and smearing based, or Hough-based [6, 11, 17]. Although projection based methods [8] have been successfully applied for machine-printed documents [10], variation in baseline and local orientation make them less effective for handwritten lines. Hough-based methods can handle documents with variation in orientation between text lines, but their performance also degrades rapidly when the baseline is not straight. Grouping based approaches use connected components(CC) to handle complex layouts, but due to proximity or touching characters across and within text lines, these methods are often inadequate.

General Terms Line detection, Arabic, Dijkstra’s shortest path algorithm, Breadth-first search, Clustering, Affinity propagation

Keywords Text line segmentation, Handwritten documents, Arabic documents ∗The partial support of this research by Defense Advanced Research Projects Agency(DARPA) through BBN/DARPA Award HR0011-08-C-0004 under subcontract 9500009235 and the US Government through NSF Award IIS-0812111 is gratefully acknowledged.

To address these problems, we adopt an approach that combines advantages of both local and global methods. We model the problem of text line extraction as a clustering problem, and propose a novel and fast way of obtaining text lines. First, we defer small components that likely correspond to accents or diacritics, to obtain an estimate of text lines based on only stable characters. This provides us two advantages. First, it speeds up the graph search and clustering method used in the next step as the running time depends quadratically on the number of components. Second, these coarse text lines can be assumed to have piece-wise constant (or smoothly varying) orientation that can be esti-

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. DAS ’10, June 9-11, 2010, Boston, MA, USA Copyright 2010 ACM 978-1-60558-773-8/10/06 ...$10.00

135

(a)

(b)

(c)

(d)

(e)

Figure 1: Sample images from our dataset. Observe that the lines have non-uniform skew across the width and character size variation is high. mated locally. We use a local orientation detection technique to obtain a sparse similarity graph containing edges between neighboring coarse components. Using a shortest path algorithm [24], we obtain a similarity between non-neighboring components which were not captured during local orientation detection. We then obtain two different estimates of text lines - one based on Breadth-first search (BFS) [25] and other based on the Affinity propagation clustering method [23]. We combine the results obtained from the two methods based on characteristics of valid text lines to obtain final coarse text lines. Finally, we assign the remaining components to corresponding text lines based on few simple rules. The proposed method has several advantages. First, given our two level approach, coarse text line estimation followed by diacritics assignment, the method is very fast. This is due to the reduction in size of similarity matrix which depends on number of components as O(n2 ). A document (Figure 1) with approximately 400 components takes less than 0.5 second to obtain coarse text lines on a P4 machine with 3 GB RAM. Second, our approach can handle both types of orientation variation problems mentioned previously. Local orientation detection can handle variation within text line and Affinity propagation based clustering handles global skew variation of text lines. Methods with only a global cost for assigning labels to each component may merge small sized text lines with bigger ones, but since we also use a local method for finding text lines our approach does not suffer from this drawback. Due to significant variability of writing styles we do not use any prior information about character shapes and sizes, and try to adapt to the properties of a given document image by using statistics obtained from the image. This makes the proposed method more robust. We achieve 96% accuracy based on pixel-matching evaluation on a dataset of 125 Arabic document images. Results on proximity analysis datasets demonstrates the effectiveness of our approach in scenarios where text lines are very closely spaced.

Figure 2: Steps for computing local orientation. First probable diacritic components are removed followed by estimation of a line in the direction of local orientation. Dots in the block represent centroids of connected components and the vertical and horizontal axes show a typical rectangular region considered for local estimation.

2. RELATED WORK Text line extraction techniques can be broadly classified as top-down projection based methods, bottom-up component grouping based methods or a hybrid of the two. Since projection based methods assume parallel text lines with sufficient gap between them, they are only effective for machine printed and a very limited class of hand printed documents. A piece-wise projection approach [1, 15] has been proposed for handwritten text lines which divides the document into vertical strips, uses horizontal projection profiles to extract components and groups these components based on heuristics for text line extraction. In [1], components are grouped by modeling the text lines as bivariate Gaussian densities and evaluating the probability of each component under each Gaussian. A similar approach [11] initially oversegments zones into text and gap regions, and uses Hidden Markov Model to find the optimal assignment of text and gap areas in each zone. The width of the zone is selected to maximize the amount of text and minimize the influence of skew in each zone. Due to large variation in the width of Arabic characters this criterion may not be always satisfied resulting in suboptimal performance of the method. Pal

Remainder of this paper is organized as follows. In Section 2, we present an overview of previous work related to handwritten text line extraction. We explain our method in detail in Section 3. A pixel-matching score based evaluation method and results of our experiments are presented in Section 4 and we conclude our paper in Section 5.

136

Figure 3: Dots are characterized by a high eccentricity value and low mass as shown in green ellipses. Main strokes having low mass has high aspect-ratio as shown in red rectangle. Accent components have low mass and aspect ratio less than two as shown in blue rectangle.

Figure 4: Local coordinate system centered at centroid of each component and divided into five regions corresponding to each possible local orientation.

et. al. [15] uses horizontal histogram of the vertical strips and the relationship of minimal values to obtain hand written text lines in Bangla document images. The piece-wise line separation computation used in their method may not work well if lines are closely spaced and orientation variation within each line is high.

and apply Breadth-first search and Affinity propagation to obtain two estimates of the coarse text lines.

3.1.1 Preprocessing We obtain all the connected components in a given document image and filter the probable diacritic components based on eccentricity, mass and size of bounding-box to obtain coarse connected components. Dots are frequently used in Arabic with main strokes and distinguish between different characters having similar shape. A high eccentricity and small mass characterizes these dots as shown in Figure 3. Accent components have low mass and aspect-ratio close to one. They are differentiated from low mass main strokes based on aspect-ratio as demonstrated in Figure 3. With this approach we were able to remove a major portion of diacritic like components, although more complete model could also take into account, the relative position of diacritic components with respect to the main character.

Connected-components based methods [22] merge neighboring connected components using heuristics based on the geometric relationship between neighboring blocks, such as distance and size compatibility. Although approaches based only on connected components are fast and can process complex layouts these methods are sensitive to topological changes of the components. Louloudis et al. [12] uses a block-based Hough transform approach to detect handwritten text lines and corrects false alarms using a merging method. A very effective and elegant method based on curve evolution and level-sets was proposed in [17, 21], but the method is very slow for high resolution images. Until recently, limited work has been done on the segmentation and recognition of handwritten Arabic text. Due to the unique nature of the script, existing methods do not always prove to be the most effective. Zahour et al. [14] used partial contour following based method to find separating lines in Arabic documents. They developed a new segmentation method [19] suited for Arabic historical manuscripts to segment the document image into three classes: text, graphics and background. Recently, in [7] a block-covering based text line segmentation method is proposed for overlapping and touching components. Faisal et al. [20] discussed preprocessing methods for handwritten Arabic and proposed a method for baseline detection of Arabic words.

3.

3.1.2 Local Orientation Detection We use a piece-wise linear approximation of orientation to obtain a sparse similarity graph with edges between components and its closest neighbors. Let the set of all coarse components be denoted by S. For each connected component Ci ∈ S, we define a local cartesian coordinate system centered at the centroid of the component and denote neighbors of Ci as N (Ci ) as given by Equation 1:

N (Ci ) = {Cj : j 6= i, Dij (x) < Rx , Dij (y) < Ry }

(1)

TEXTLINE SEGMENTATION

Our text line extraction method consists of two steps - coarse text line estimation and diacritic assignment. Figure 2 shows the steps involved and each is explained in detail in following subsections.

Dij (x) = kCj (x) − Ci (x)k,

Dij (y) = kCj (y) − Ci (y)k (2)

where Dij (x) and Dij (y) represents the horizontal and vertical distances between Cj and Ci respectively. Ci (x) and Ci (y) denote the x and y coordinates of centroid of Ci . Rx and Ry are determined from the statistics of components in S and the size of current component Ci as follows : Let Hmed and Wmed be the median height and width of bounding-box of all components in S and Wcur be the width of current component. The initial Rx and Ry is given by Equation 3:

3.1 Coarse Textline Estimation First, we remove probable diacritic and accent components to obtain only those components which resemble the main strokes of an Arabic character. We then, use a local orientation detection method and Shortest path algorithm to obtain a similarity measure between primary components

137

Rx = Wcur + t1 ∗ Wmed ,

Ry = t2 ∗ Hmed

(3)

where t1 and t2 are parameters whose value depends on the average character width and gap between the characters of a script. If the number of neighboring components is not sufficient for local orientation estimation, we increase the dimension of region in steps by a factor as given by Equation 4:

Rx

new

= f1 ∗ Rx

old

,

Ry

new

= f2 ∗ Ry

old

(a)

Figure 5: Two types of similarities computed using (a) direct distance to locally estimated line (b) sum of the distances along the shortest path.

(4)

Algorithm 1 DiactricsAssignment(D,C,T) Require: D = {di , i := 1 : m} {Diacritic components} Require: C = {ci , i := 1 : n} {Coarse components} Require: T = {ti , i := 1 : k} {Estimated coarse text lines} for i = 1 to m do Find S = {cj ∈ C s.t. di lies inside the BB of cj } if S 6= Φ then Find cj ∈ S whose orientation line is closest to centroid of di Assign di to tk where cj ∈ tk Continue with next i ← i + 1 end if Find S = {cj ∈ C s.t. BB of di overlaps with BB of cj } if S 6= Φ then Find cj ∈ S whose orientation line is closest to centroid of di Assign di to tk where cj ∈ tk Continue with next i ← i + 1 end if Find S = {cj ∈ C s.t. distance of di to cj < Hmedian } if S 6= Φ then Find cj ∈ S whose orientation line is closest to centroid of di Assign di to tk where cj ∈ tk Continue with next i ← i + 1 end if end for

We divide the coordinate system into eight sectors around the origin as shown in Figure 4 and pair the diagonally adjacent sectors (shown by numbers 1-4) to quantize the orientation of text line at Ci . We also define a focused region at an angle of ±10 degrees with x-axis in each of the four quadrants (region 5 in Figure 4). We consider all the neighboring components defined by Equation 1 which lie in the rectangular region centered at the centroid of current CC. The dimensions of this region given by Equation 3 are adaptive based on the size of current CC and median height and width of all coarse components. We obtain the count of neighboring components in each of the five regions and find the region Ri max given by Equation 5 with maximum components.

Ri max = max{Count(Rj , Ci )} j

(b)

(5)

where Count(Rj , Ci ) denotes the count of components in region Rj . We estimate the direction by obtaining the least square estimate of a line passing through Ci using the centroid of components in region Ri max . The orientation of line determines local orientation at Ci . We then find the distance between the centroid of each neighboring component in region Ri max and the estimated line to compute the similarity between the current component Ci and neighboring components using Equations 6 and 7.

3.1.3 Text Lines Extraction using Breadth-First Search |yj − mxj − b| √ Dist(Cj , Ci ) = Dist(Cj , Li ) = m2 + 1

Next we find all disjoint subsets of vertices of the similarity graph for which there exist a path from each element to every other element in the set. These subsets represent connected-components of the similarity graph which are obtained using Breadth-First search (BFS) [25]. Ideally, if all the local estimations are correct then each connected component represents a text line. In practice, the variation in the size of characters and proximity between lines causes errors in the orientation estimation. For example, while computing the locally oriented neighbors of a component, if any component Ci of another line is also present in that region, then it will also have an edge to the current component. Hence, instead of making a hard decision on the assignment of a component to a text line, we also compute soft similarities between components and use a clustering method called Affinity propagation to find the optimal assignment in the next step. The running time of BFS is O(|V | + |E|), where |V| is the number of nodes and |E| is the number of edges

(6)

where m is the slope of line and b=0 is the y-intercept of estimated line Li at component Ci . In parallel, we build a similarity graph in which we put an edge between current component and neighbors in region Ri max with similarity value given by Equation 7:

S(Cj , Ci ) = exp (−Dist(Cj , Li ))

(7)

When Dist(Ci , Cj ) is not equal to Dist(Cj , Ci ), we retain the minimum of the two so that the similarity matrix is symmetric. This will frequently happen when local orientation estimation is done at Cj and Ci ∈ Rj max .

138

in the graph. Since the graph in our case is sparse, running time is almost linear (O(|V |)) in the number of coarse components.

from candidate exemplar point k to point i, shows the accumulated evidence for how appropriate it would be for point i to choose point k as its exemplar. The self-availability is updated in a different way as given by Equation 12.

r(i, k) ← s(i, k) − ′max {a(i, k′ ) + s(i, k′ )} ′ k ,k 6=k

(a)

(b)

(c)

(d) a(i, k) ← min{0, r(k, k) −

Figure 6: Examples of diacritics with bounding-box fully inside the bounding-box of a primary component.

3.1.4 Textline Extraction using Affinity Propagation

X

(a) Dist(Cl , Ck )

max{0, r(i′ , k)}} (11)

i′ ,i′ ∈{k,i} /

a(k, k) ← max{0,

Before applying the Affinity propagation(AP) clustering method [23], we first find similarities between all components in the graph as follows. For each node in the local orientation graph we find the shortest path to every other node using Dijkstra’s shortest path algorithm [24]. The distance between two nodes is then based on sum of the distances in shortest path as given by Equation 8:

DistSP (Ci , Cj ) =

X

(10)

X

r(i′ , k)}

(12)

i′ ,i′ 6=k

(b)

(c)

(8)

Cl ,Ck ∈SP

Figure 7: Examples of diacritics with bounding-box overlapping in vertical direction with bounding-box of a primary component.

where SP is the set of nodes in shortest path from Ci to Cj . Once the distance is obtained, the similarity between node i and node j is computed using Equation 7. The running time of this algorithm is O(|V |2 +|E|), but for sparse graphs it can be implemented more efficiently in O(|E| + |V |log|V |) using an adjacency list and Fibonacci heap as a data structure. We also assign a similarity to each component and its neighbors N (Ci ) ∈ / Rimax , by finding the distance to the estimated line at Ci as given by Equation 9, where distance Dist(Cj , Li ) is given by Equation 6. α > 1 is a penalizing factor for the component since it does not belong to Ri max . Hence, in this way we have a non-zero probability for each component Cj ∈ N (Ci ) not detected locally, to become associated with the same text line as Ci on the basis of AP.

S(Cj , Ci ) = exp (−α ∗ Dist(Cj , Li ))

(a)

(b)

(c)

Figure 8: Example of diacritic components which do not overlap with bounding-box of any primary component.

(9)

Affinity Propagation does not require an initial estimate of the number of clusters. It takes as input, a similarity matrix of real-valued entries, where the similarity s(i,k) indicates how well the data point with index k is suited to be the exemplar for data point i. The diagonal entries of the similarity matrix representing self-similarities are called preferences. This value for each data point k encodes the likelihood of that point to be selected as exemplar. The number of clusters is influenced by the values of the preferences, but also emerges from the message-passing procedure. There are two kinds of messages exchanged between data points. The responsibility r(i,k) given by Equation 10, sent from data point i to point k, represents the evidence for how well-suited point k is to serve as the exemplar for point i. The availability a(i,k) given by Equation 11, sent

(a)

(b)

Figure 9: (a) Conflict in case of vertical overlap based diacritic assignment (b) Conflict is resolved based on the distance between centroid of diacritic and orientation line.

139

(a)

(b)

(c)

Figure 10: (a) Split error (b) Merge error (c) Mixed error

3.1.5 Error Detection and Correction Errors in text line extraction obtained using BFS are typically a result of either a missing edge between two adjacent components or due to an edge between components of different text lines in the initial orientation graph. The former will occur when the local orientation detection fails to identify one component as a neighbor of other and vice versa. This results in two or more text lines in result, corresponding to a single text line in ground truth (Figure 10(a)). Since the neighboring components will assume a smaller distance to the orientation line of a component, these adjacent components will have high similarities and hence are likely to get clustered together using AP. One positive thing is this error preserves the reading order of text lines and hence for applications which consider reading order of text lines, correction is trivial. We correlate each text line in the results obtained from AP to detect disjoint multiple regions from BFS estimate, to detect such errors.

Figure 11: Sum of local projection profiles at each primary component in the estimated direction for incorrect and correct scenarios.

The latter occurs if the least-square estimate of orientation involves some component of other text line, then the two text lines may grouped together to form a single text line. Our adaptive region size computation at different scales for local orientation estimation minimizes this error as demonstrated by our results on proximity data sets. The errors produced in this case are very similar in structure to those produced by touching of characters, as observed in results for our dataset. Figure 10(b,c) shows possible error scenarios in our approach. As shown, errors of this type can be detected if multiple lines from one result have considerable overlap with single text line in another. Local projection profiles in the direction of estimated orientation of text line at each component can also be used to discriminate between a correct and an incorrect case. Figure 11 shows the sum of local projection profiles over the text line are very different for the two cases.

is selected. Finally, remaining diactric components are assigned based on minimum distance to the orientation line of coarse components.

3.2 Diacritics Assignment

Figure 12: (a) A typical ground-truth zone with rectangular boxes. Each box has the same line-id. (b) Corresponding result zone from our method.

In the next step, components which were removed in the first step are assigned to text lines. For each diacritic component, we find the best matching coarse component, and assign the label of coarse component to it. Algorithm 1 provides the sequence of steps used in our approach. First preference is given to the coarse components whose bounding-box fully contain the bounding-box of diacritic component. Figure 6 shows some examples. Second preference is given to those coarse components whose bounding-box overlaps with the bounding-box of a diacritic component as shown in Figure 7. If there are multiple candidates in any of the above cases, it is resolved on the basis of centroid distance of diacritics to the orientation line computed at coarse components as depicted in Figure 9. Component with minimum distance

4. EXPERIMENTS AND ANALYSIS We evaluate our results based on percentage of foreground pixels matched between a ground truth text line and a result text line. While ground truth text lines are represented as rectangular boxes for words in the text line, result of our method is a rectangular region around the text line (Figure 12). We refer to these rectangles as zones. Ground-truth zones having same line-id are merged to form a single zone for evaluation. We define the matching score metric between a result zone ri and ground truth zone gj by Equation 13:

140

Proximity Analysis (Absolute )

Proximity Analysis (Relative)

100

100 :90

MS

:90

MSthres:95

MS

:95

MS

thres

thres

80

60

F1−Score

F1−Score

80

40

20

0 0

thres

60

40

20

20

40 60 Number of pixels moved

80

0 0

100

(a)

0.5 1 1.5 Fraction(r) of avgerage text line spacing moved

(b)

Figure 13: Plot of F1-Score on (a) Absolute proximity datasets (b) Relative proximity datasets

Table 1: Results obtained on a dataset of 125 images with matching score threshold as 90% and 95% M Sthres 90 95

#GT 1974 1974

#Result 1943 1943

Precision 96.35% 91.61%

M atchingScore(ri , gj ) =

Recall 94.83% 90.17%

T (P (ri ) ∩ P (gj )) T (P (ri ) ∪ P (gj ))

F1-Score 95.6 90.9

(13)

where P(s) represents the foreground pixels in zone s and T is an operator which counts the number of pixels in a zone. Using this metric we obtain a matching score table between result zones and ground truth zones. If a result zone overlaps with a ground truth zone and has a matching score greater than some predefined threshold (95% in our case), we declare it as matched. Result zones which are not matched to any ground truth zones are counted as False positives (FP) and the ones which are matched are counted as True positives (TP). Similarly, ground truth zones which do not match to any result zones are declared as False negatives (FN). Based on these quantities we compute precision, recall and F1 -Score as given by Equation 14 and Equation 15 respectively:

P recision =

TP TP + FP

F1 − Score =

Recall =

TP TP + FN

2 ∗ P recision ∗ Recall P recision + Recall

our method on this dataset for two different matching score thresholds. To analyze the robustness of our method, we generated a proximity dataset from these 125 images by moving each text line closer to the text line above it. Two types of proximity datasets were generated. In the first type, which we call Absolute proximity, each text line except the top-most one, was moved closer to the line above it, in multiples of 10 pixels to obtain 10 datasets each of 125 images. Figure 13(a) shows plot of F1 -Score obtained using our method. Another set of datasets were created based on movement of text lines in steps of some fixed fraction of average distance between consecutive lines. For each text line we compute the mean of ordinates (vertical axis coordinates) of left-top vertex of each component to obtain Ytop . Similarly, mean of ordinates of right-bottom vertex of components in the text line above, is computed to obtain Ybot . The average distance is the difference between Ybot and Ytop . Number of pixel spacings moved vertically is adaptively decided by a fixed percentage of this average distance and the new spacing between text lines in each step is determined by Equation 16:

Dnew = (1 − r) ∗ Davg

r = {0.1, 0.2, ...m}

(16)

where Davg is the average spacing computed and Dnew is the new spacing between the text lines. Figure 13(b) shows plot of F1 -Score obtained. Since the parameters used in the proposed method is based on the statistics of given image, method works well even for low resolution images. These artificial datasets along with the original dataset is available for download at [18].

(14)

5. CONCLUSIONS AND FUTURE WORK

(15)

We presented a novel method for segmenting text lines in handwritten Arabic document images. Our main contribution lies in the approach for computing similarity between text components based on local orientation detection and shortest path in graphs. We used these similarities to cluster the text components using Affinity propagation which auto-

Our dataset consists of 125 handwritten Arabic document images with total 1974 ground truth text lines. Figure 1 shows some of the sample images from our dataset. Size of each image is 5100x6600. Table 1 shows the result of

141

matically finds the appropriate number of clusters based on a globally defined energy measure. Hence, there is no need of specifying any estimate of number of text lines. We also proposed a new approach for detecting and correcting segmentation errors based on the two different results obtained to achieve high accuracy. The main advantage of our method is that textlines with both uniform and non-uniform skew can be segmented properly. Moreover, text lines with extremely less spacing can be segmented accurately as demonstrated by our results on proximity datasets. Our method is fast due to a two-level approach to final text line extraction. By reducing the number of components in Affinity propagation and graph-search method, we achieve a competitive time performance. At the same time, we believe that more work is required in diacritics removal and correction of touchingerrors. Our future extension of this work is to address these problems.

6.

image segmentation into textlines and words,” Pattern Recognition, Volume 43, Issue 1, January 2010, Pages 369-377 [12] G. Louloudis, B. Gatos, C. Halatsis, “Text Line Detection in Unconstrained Handwritten Documents Using a Block-Based Hough Transform Approach,” ICDAR, vol. 2, pp.599-603, Ninth International Conference on Document Analysis and Recognition (ICDAR 2007) Vol 2, 2007 [13] Amin A. “Off-line Arabic character recognition : The state of the art”, Pattern Recognition, Vol. 31, pp. 517-530, 1998. [14] A. Zahour, B. Taconet, P. Mercy, S. Ramdane, “Arabic handwritten text-line extraction,” In:Proceedings of the Sixth International Conference on Document Analysis and Recognition, 2001, pp. 281-285. [15] U. Pal, S. Datta, “Segmentation of Bangla unconstrained handwritten text,” In:Proc. of the Seventh International Conference on Document Analysis and Recognition, vol.2, 2003, pp. 1128–1132 [16] Venu Govindaraju, Huaigu Cao and Anurag Bhardwaj, “Handwritten Document Retrieval Strategies”, Proc. of ICDAR worskhop on Noisy Text Analytics (AND), Spain, 2009. [17] Yi Li, Yefeng Zheng and David Doermann, “Detecting Text Line in Handwritten Documents,” ICPR’06, pages 1030–1033, 2006. [18] Handwritten Arabic Proximity Datasets. Language and Media Processing Laboratory. http://lampsrv02.umiacs.umd.edu/projdb/project.php [19] W. Boussellaa, A. Zahour, B. Taconet, A.Benabdelhafid, A. Alimi, “Segmentation texte /graphique : Application au manuscrits Arabes Anciens.”, Neuvi`eme Colloque International ˇ Francophone sur lSEcrit et le Document, Fribourg, Suisse, 18-21 Septembre 2006,pp. 139-144 [20] F. Farooq, V. Govindaraju, and M. Perrone, “Preprocessing Methods for Handwritten Arabic Documents”, Proc. Int’l Conf. Document Analysis and Recognition, pp. 267-271, 2005. [21] Du, X., Pan, W. et Bui, T. D., “Text line segmentation in handwritten documents using mumford-shah model,” Pattern Recogn., 42(12):3136-3145, 2009. [22] U. V. Martin and H. Bunke., “Text line segmentation and word recognition in a system for general writer independent handwriting recognition,” In Proc. Intl. Conf. on Document Analysis and Recognition, pages 159-163, 2001. [23] Brendan J. Frey and Delbert Dueck, “Clustering by Passing Messages Between Data Points,” Science 315, 972-976 [24] Cormen, Thomas H.; Leiserson, Charles E.; Rivest, Ronald L.; Stein, Clifford (2001). “Section 24.3: Dijkstra’s algorithm,” Introduction to Algorithms (Second ed.). MIT Press and McGraw-Hill. [25] Cormen, Thomas H.; Leiserson, Charles E.; Rivest, Ronald L.; Stein, Clifford (2001). “Section 22.2: Breadth First Search,” Introduction to Algorithms (Second ed.). MIT Press and McGraw-Hill.

REFERENCES

[1] Manivannan Arivazhagan, Harish Srinivasan, and Sargur Srihari, “A statistical approach to line segmentation in handwritten documents,” Volume 6500. SPIE, 2007. [2] Masaki Yamaoka and Osamu Iwaki, “Document layout analysis using pattern classification method,” Lecture Notes in Computer Science, Vol. 1024/1995, pp. 524-525 [3] Chih-Hong Kao, Hon-Son Don, “Skew Detection of Document Images Using Line Structural Information,” icita, vol. 1, pp.704-715, Third International Conference on Information Technology and Applications (ICITA’05) Volume 1, 2005 [4] Arvind K. R., Jayant Kumar and Ramakrishnan A. G.,“Entropy Based Skew Correction of Document Images,”Lecture Notes in Computer Science, Vol. 4815/2007, Springer, pp. 495-502, 2007 [5] U.-V. Marti, H. Bunke, “Text Line Segmentation and Word Recognition in a System for General Writer Independent Handwriting Recognition,” pp.0159, Sixth International Conference on Document Analysis and Recognition (ICDAR’01), 2001 [6] Z. Razak, K. Zulkiflee, “Off-line Handwriting textline segmentation: a review,” International Journal of Computer Science and Network Security 8(7)(2008) 12–20. [7] Zahour A.,Taconet B., Likforman-Sulem L., Bousella W., Overlapping and multi-touching text-line segmentation by block covering analysis, Pattern Analysis & Applications, DOI 10.1007/s10044-008-0127-9, July 2008. [8] B. Yanikoglu, P. A. Sandon, “Segmentation of off-line cursive handwriting using linear programming”, Pattern Recognition 31(12)(1998) 1825-1833. [9] R´ejean Plamondon, Sargur N. Srihari, “On-Line and Off-Line Handwriting Recognition: A Comprehensive Survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 22, no. 1, pp. 63–84, Jan. 2000 [10] Y. Lu, “Machine printed character segmentation: an overview,” Pattern Recognition 28, 67-80 (1995). [11] Vassilis Papavassiliou, Themos Stafylakis, Vassilis Katsouros, George Carayannis, “Handwritten document

142