A simple text ine segmentation method for handwritten documents1

A simple text-line segmentation method
for handwritten documents
M.Ravi Kumar1, R. Pradeep2, B.S.Puneeth Kumar3, and Prasad Babu4
Dept. of P.G.Studies in Computer Science, Kuvempu University
Jnana Sahyadri, Shankaraghatta-577451
E-mail: ravi2142@yahoo.co.in, punithashp@gmail.com, kamitkar.prasad9@gmail.com

Abstract:
Text line segmentation is an important step mainly due to the irregularity of layout and character shapes
because inaccurately segmented text lines will cause errors in originated from the variability of writing styles. For
the recognition stage.. The nature of handwriting makes the unconstrained handwritten documents, text line segmentation
process of text line segmentation very challenging. Text and character segmentation-recognition are not solved though
characteristics can vary in font, size, orientation, alignment, enormous efforts have been devoted to them and great
color, contrast, and background information. These variations advances have been made. Text line segmentation of
turn the process of word detection complex and difficult. handwritten documents is much more difficult than that of
Since handwritten text can vary greatly depending on the user printed documents. Unlike that printed documents have
skills, disposition and cultural background. In this work we approximately straight and parallel text lines, the lines in
have proposed the method which works on the different handwritten documents are often un-uniformly skewed and
intensity values for extracting the text-lines. curved. Moreover, the spaces between handwritten text lines
are often not obvious compared to the spaces between within-
line characters, and some text lines may interfere with each
Keywords: other. Therefore, many text line detection techniques, such as
projection analysis and K-nearest neighbor connected
Handwritten document, segmentation, text-lines; components (CCs) grouping, are not able to segment
handwritten text lines successfully[2].
1. Introduction: On the other hand, piece-wise projections are sensitive to
character’s size variation within text lines and significant gaps
Segmentation of a document image into between successive words. These occurrences influence the
its basic entities, namely, text lines and words, is considered as effectiveness of smearing methods too. In such cases, the
a non-trivial problem to solve in the field of handwritten results of two adjacent zones may be ambiguous, affecting the
document recognition. The difficulties that arise in drawing of text-line separators along the document width. To
handwritten documents make the segmentation procedure a deal with these problems we introduce a smooth version of the
challenging task. Different types of difficulties are projection profiles to over segment each zone into candidate
encountered in the text line segmentation and the word text and gap regions. Then, we reclassify these regions by
segmentation procedure. In case of text line segmentation applying an HMM formulation that enhances statistics from
procedure, major difficulties include the difference in the skew the whole document page. Starting from left and moving to the
angle between lines on the page or even along the same text right we combine separators of consecutive zones considering
line, overlapping words and adjacent text lines touching. their proximity and the local foreground density [3].
Furthermore, the frequent appearance of accents in many These piece-wise projection based methods have a few
languages (e.g. French, Greek) makes the text line shortcomings: (a) they generate too many potential separating
segmentation a challenging task. In word segmentation, lines, (b) the parameter of stripe width is predefined, (c) text-
difficulties that arise include the appearance of skew and slant lines should not have significant skew as mentioned in and (d)
in the text line, the existence of punctuation marks along the if there is no potential piece-wise line in the first and last
text line and the non- uniform spacing of words which is a stripes, drawing a complete separating line will become
common residual in handwritten documents[1]. Text line impossible in the any algorithms. These shortcomings are
segmentation from document images is one of the major observed based on an experiment that we have conducted with
problems in document image analysis. It provides crucial a number of text-pages. Some authors also used skew
information for the tasks of text block segmentation, character information for text-line separation. In an unconstrained
seg-mentation and recognition, and text string recognition. handwritten text-page, it is very difficult to detect the
Whereas the difficulty of machine-printed document analysis orientation of each line on the basis of the skew calculated for
mainly lies in the complex layout structure and degraded the entire page. Therefore, these methods may not work
image quality, handwritten document analysis is difficult properly [4].

1

The concept of the Hough transform is employed in the field densities to assign overlapping CCs into text lines [7].
of document analysis for many purposes such as skew Experimental results on a collection of 720 documents
detection, line detection, slant detection and text-line (English, Arabic and children's handwriting) show that
segmentation. The Hough transform is employed for text-line 97.31% of text lines were segmented correctly. The writers
segmentation in different scripts. A block-based Hough mention that “a more intelligent approach to cut an
transform is presented which is a modification of the overlapping component is the goal of future work”. A recent
conventional Hough transform methodology. The algorithm approach [8] uses block-based Hough transform to detect lines
includes partitioning of the connected component domain into and merging methods to correct false alarms. Although the
three spatial sub-domains and applying a block-based Hough algorithm achieves a 93.1% detection rate and a 96%
transform to detect the potential text lines [5]. recognition rate, it is not flexible to follow variation of skew
Many efforts have been devoted to the difficult problem of angle along the same text line and not very precise in the
handwritten text line segmentation. The methods can be assignment of accents to text lines. Li et al. [9] discuss the
roughly categorized into three classes: top-down, bottom-up, text-line detection task as an image segmentation problem.
and hybrid. Top-down methods partition the document image They use a Gaussian window to convert a binary image into a
recursively into text regions, text lines, and words/characters smooth gray-scale. Then they adopt the level set method to
with the assumption of straight lines. Bottom-up methods evolve text-line boundaries and finally, geometrical constrains
group small units of image (pixels, CCs, characters, words, are imposed to group CCs or segments as text lines. They
etc.) into text lines and then text regions. Bottom-up grouping report pixel-level hit rates varying from 92% to 98% on
can be viewed as a clustering process, which aggregates image different scripts and mention that “the major failures happen
components according to proximity and does not rely on the because two neighboring text lines touch each other
assumption of straight lines. Hybrid methods combine bottom- significantly”. A similar method [10] evaluates eight different
up grouping and top-down partitioning in different ways. All spatial measures between pairs of CCs to locate words in
the three approaches have their disadvantages. Top-down handwritten postal ad- dresses. The best metric proved to be
methods do not perform well on curved and overlapping text the one which combines the result of the minimum run-length
lines. The performance of bottom-up grouping relies on some method and the vertical overlapping of two successive CCs.
heuristic rules or artificial parameters, such as the between- Additionally, this metric is adjusted by utilizing the results of
component distance metric for clustering. On the other hand, a punctuation detection algorithm (periods and commas).
hybrid methods are complicated in computation, and the Then, a suitable threshold is computed by an iterative
design of a robust combination scheme is non-trivial [2]. procedure. The algorithm tested on 1000 address images and
A thinning operation has also been used by other researchers performed an error rate of about 10%.
for text-line segmentation of Japanese and Indian text Manmatha and Rothfeder [11] propose an effective for noisy
documents. Thinning algorithms followed by post-processing historical documents scale space approach. The line image is
operations are employed for the entire background region of filtered with an anisotropic Laplacian at several scales in order
an input text image to detect the separating borderlines. to produce blobs which correspond to portions of characters at
Recently, some techniques have used level set, active contour small scales and to words at larger scales. The optimum scale
and a variational Bayes method for text-line segmentation. is estimated by three different techniques (line height, page
Density estimation and the level set method (LSM) were averaging and free search) from which the line height showed
utilized for text-line segmentation. A probability map is best results. Much more challenging task is line segmentation
estimated from an input document image, where each element in historical documents due to a great deal of noise. Feldbach
represents the probability of the original pixel belonging to a and Tonnies [12] have proposed a bottom up method for
text line. The level set method (LSM) is utilized to determine historical church documents that requires parameters to be set
the boundary evolution of neighboring text lines. At first, a according to the type of handwriting. They report a 90%
matched filter bank approach is used for smoothing the input correct segmentation rate for constant parameter values which
text image. The central line of text-line components is then rises to 97% for adjusted ones.
computed using ridges over the smoothed image. Finally, the Another integrated system for such documents [13] creates a
active contours (snakes) over the ridges are adapted to obtain foreground/background transition count map to find probable
the text-line segmentation result [4]. locations of text lines and applies min-cut/max-flow algorithm
to separate initially connected text lines .The method performs
2. Related Work: high accuracy (over 98%) in 20 images of George
Washington's manuscript.
In this section, we give a brief review of recent work on text
line and word segmentation in handwritten document images.
As far as we know, the following techniques either achieved 3. Segmentation challenges:
the best results in the corresponding test datasets, or are
elements of integrated systems for specific tasks. One of the In this section we the challenges involved in the segmentation
most accurate methods uses piece-wise projection profiles to of the text-lines. When dealing with handwritten text, line
obtain an initial set of candidate lines and bivariate Gaussian

2

segmentation has to solve some obstacles that are uncommon 4. Motivation:
in modern printed text [6]. Among the most predominant are: The above challenges motivated us to take this challenging
work.
3.1 Skewed lines: Lines of text in general are not straight.
These lines are not parallel to each other. 5. Proposed Method:

In the proposed method our objective is to identify the
boundary of the text-lines, which consist two steps (i)
Generating partial boundary line and (ii) Generating complete
boundary line. By using these partial lines it is very difficult to
differentiate between the two lines. Because these partial lines
are having gaps and they are broken. Thus we need to
construct the complete line which acts as differentiator for
identifying the text-lines. In constructing the complete line we
will consider the highest frequent y co-ordinate value in each
partial boundary lines, y co-ordinate values are the complete
Figure 3.1: Skewed Lines
boundaries for each text-lines. In last step we segment the
each text-line by representing the each text-line with different
3.2 Fluctuating lines: Lines of text are partially or fully
colors.
connected to other text-lines.

5.1 Generating partial text-line boundary:

The partial text-line boundary is generated by blocking the
text-lines. Using morphological operations like erosion. Here
blocking is done for filling the holes and gaps between the
words. This helps in drawing the partial line. Here the lines
are drawn at the edge of the every blocked text-line. However
these partial lines are not sufficient to differentiate the text-
lines. This is shown in figure 5.1.
Figure 3.2: Fluctuating Lines
5.2 Generating complete text-line boundary:
3.3 Line proximity: Small gaps between neighboring text
lines will cause touching and overlapping of components, As in the previous step we obtained the partial boundary lines,
usually words or letters, between lines and irregularity in these broken boundary lines are not sufficient for segmenting
geometrical properties of the line, such as line width, height, the text-lines, thus we need to generate complete line which is
distance in between words and lines, leftmost position etc. continuous. So now by using these partial lines, we draw the
complete boundary lines with the help of the frequent vertical
points. These y co-ordinate values are used to differentiate
between text-lines, and these complete boundary lines helps in
segmentation. This is showed in figure 5.2

5.3 Text-line segmentation:

Once the complete boundary lines were drawn, it is easy to
segment the text-lines by assigning different values to each
character in between the lines. In this section, we give the
different intensity values for the characters in a text-line
between the two drawn lines. Through the different intensity
values of the characters in a text-line, this method recognizes
Figure 3.3: Line proximity the different lines. The segmentation for different languages
Source: [6] are showed in figure 5.3

3

Figure 5.1: Illustration of Partial text-line boundary detection. (a) A handwritten document. (b) Blocked text-
lines. (c) Partial boundary lines

Figure 5.2: Illustration of complete text-line boundary detection. (a) Complete boundary lines with partial
boundary lines. (b) Complete boundary lines without partial boundary lines

4

Figure 5.3: Illustration of segmented text-lines. (a) Segmented English document. (b) Segmented Kannada
document (c) Segmented Hindi document (d) Segmented6. Block diagram of proposed method:
English document

5

7. Results and discussion:
We conducted the experiments on the languages like Kannada, [7] Z. Razak, K. Zulkiflee, et al., Off-line handwriting text line
English, Hindi and Arabic of 500 documents (kannada-150, segmentation: a review, International Journal of Computer
English-200, Hindi-100 and Arabic-50) and obtained the Science and Network Security 8 (7) (2008) 12–20.
accurate results. [8] G. Louloudis, B. Gatos, C. Halatsis, Text line detection in
unconstrained handwritten documents using a block-based
Hough transform approach, in: Proceedings of International
8. Conclusion and future work: Conference on Document Analysis and Recognition, 2007, pp.
599–603
In this work we are mainly concentrated on extracting required [9] Y. Li, Y. Zheng, D. Doermann, S. Jaeger, Script-
text-line from the given document and obtained the specified independent text line segmentation in freestyle handwritten
text-line accurately. The proposed method works well for documents, IEEE Transactions on Pattern Analysis and
segmenting the text-line of handwritten document. Our Machine Intelligence 30 (8) (2008) 1313–1329.
method only works on the fixed word length and text-lines [10] G. Seni, E. Cohen, External word segmentation of off-
without skew; in our future work we will improve the results line handwritten text lines, Pattern Recognition 27 (1994) 41–
by segmenting the text-lines with above described limitations. 52.
[11] R. Manmatha, J.L. Rothfeder, A scale space approach for
9. References: automatically segmenting words from historical handwritten
documents, IEEE Transactions on Pattern Analysis and
[1] G. Louloudisa, B.Gatosb, I.Pratikakisb, C.HalatsisaText Machine Intelligence 27 (8) (2005) 1212–1225.
line and word segmentation of handwritten documents. Pattern [12] M. Feldbach, K.D. Tonnies, Line detection and
Recognition (2008) pp. 3169 – 3183 segmentation in historical church registers, in: Proceedings of
[2] Fei Yin, Cheng-LinLiu. Handwritten Chinese text line International Conference on Document Analysis and
segmentation by clustering with distance metric learning Recognition, 2001, pp. 743–747.
Pattern Recognition (2009) pp. 3146 -- 3157 [13] D.J. Kennard, W.A. Barrett, Separating lines of text in
[3] Vassilis Papavassiliou, Themos Stafylakis, et al.., free-form handwritten historical documents, in: Proceedings of
Handwritten document image segmentation into textlines and International Workshop on Document Image Analysis for
words. Pattern Recognition (2010) pp. 369 -- 377 Libraries, 2006, pp. 12–23.
[4] Alireza Alaei UmapadaPal, et al..,. A new scheme for
unconstrained handwritten text-line segmentation. Pattern
Recognition (2011) pp. 917–928
[5] A Block-Based Hough Transform Mapping for Text Line
Detection in Handwritten Documents G. Louloudis1, B.
Gatos2, I. Pratikakis2, K. Halatsis1
[6] Ashu Kumar , Simpel Rani Jindal, Galaxy Singla Line
segmentation using contour tracing, 2012 Vol 3

6

A simple text ine segmentation method for handwritten documents1

More Related Content

Viewers also liked (16)

Similar to A simple text ine segmentation method for handwritten documents1 (20)

Recently uploaded (20)

A simple text ine segmentation method for handwritten documents1