Enhancement and Segmentation of Historical Records

David C. Wyld et al. (Eds) : ACITY, DPPR, VLSI, WiMNET, AIAA, CNDC - 2015
pp. 95–113, 2015. © CS & IT-CSCP 2015 DOI : 10.5121/csit.2015.51309
ENHANCEMENT AND SEGMENTATION
OF HISTORICAL RECORDS
Soumya A1
and G Hemantha Kumar2
1
Dept. of Computer Science & Engg,
R V College of Engineering, Bangalore, India
soumyaa@rvce.edu.in
2
Dept. of Studies in Computer Science,
University of Mysore, Mysore, India
ghk.2007@yahoo.com
ABSTRACT
Document Analysis and Recognition (DAR) aims to extract automatically the information in the
document and also addresses to human comprehension. The automatic processing of degraded
historical documents are applications of document image analysis field which is confronted with
many difficulties due to the storage condition and the complexity of the script. The main interest
of enhancement of historical documents is to remove undesirable statistics that appear in the
background and highlight the foreground, so as to enable automatic recognition of documents
with high accuracy. This paper addresses pre-processing and segmentation of ancient scripts,
as an initial step to automate the task of an epigraphist in reading and deciphering inscriptions.
Pre-processing involves, enhancement of degraded ancient document images which is achieved
through four different Spatial filtering methods for smoothing or sharpening namely Median,
Gaussian blur, Mean and Bilateral filter, with different mask sizes. This is followed by
binarization of the enhanced image to highlight the foreground information, using Otsu
thresholding algorithm. In the second phase Segmentation is carried out using Drop Fall and
WaterReservoir approaches, to obtain sampled characters, which can be used in later stages of
OCR. The system showed good results when tested on the nearly 150 samples of varying
degraded epigraphic images and works well giving better enhanced output for, 4x4 mask size
for Median filter, 2x2 mask size for Gaussian blur, 4x4 mask size for Mean and Bilateral filter.
The system can effectively sample characters from enhanced images, giving a segmentation rate
of 85%-90% for Drop Fall and 85%-90% for Water Reservoir techniques respectively.
KEYWORDS
Document Analysis, Preprocessing, Filters, Segmentation, Drop Fall Technique, Water
Reservoir Technique
1. INTRODUCTION
A generic Optical Character Recognition (OCR) system comprises of different stages like
preprocessing, segmentation, feature extraction and classification. Preprocessing is one of the
most interesting and challenging topics in DAR. Preprocessing of document involves converting
scanned images or photographed images of machine printed or handwritten text which may

96 Computer Science & Information Technology (CS & IT)
include numbers, letters and symbols into system processable format. Segmentation is an
important assignment of any OCR system and it separates the image text documents into
lines,words and characters. Hence the accuracy of OCR system primarily depends on the
segmentation algorithm been used.
Segmentation of handwritten text of Indian languages is challenging when compared with Latin
based languages because of its structural complication and presence of compound characters.
This complexity increases further if were to recognize text of ancient Indian or non-Indian
epigraphical documents. The epigraphical records engraved on stones, rocks, pillars or on some
other writing material are non-linear in their shapes and non-uniform in their sizes.
Raw image of an epigraph contains unwanted symbols or marks, noise embedded and text
engraved with much skew.The spacing between characters and also between the lines and the
skew could complicate the process of translating the scripts. Some touching lines as well as
characters complicates the process of segmentation which is input for the recognition process in
the later stages. Hence the input document image of epigraphs is to be preprocessed for removal
of noise, skew detection and correction, followed by segmentation of characters [1].
Inspite of several positive works on OCR across the world, development of OCR tools in Indian
languages is still a challenging task. Character segmentation plays an important role in character
recognition since incorrectly segmented characters are susceptible to be recognized wrongly.
Hence the proposed work focuses on preprocessing and segmentation of ancient handwritten
documents. This is an initial step towards developing OCR for ancient scripts, which can be used
by archaeologists and historians for digitization and further exploration of ancient records.
This paper is organized as follows: Section 2 elaborates the related works in the field. The system
architecture is highlighted Section 3. The theory and related mathematical background of the
approaches in current system is discussed in Section 4. Methodology is given in Section 5.
Experimental results and performance analysis is covered in Section 6 and Section 7 provides
conclusion.
2. RELATED WORK
Researchers have worked on many approaches for preprocessing and segmentation of various
languages. In this section, some of the works are discussed.
The linear Unsharp Masking (USM) technique [2] is adopted to increase the pictorial presence of
an image by highlighting its regularity contents to improve the edge and detailed information in
it. Nevertheless this method is easy and gives good result for many applications. It has two
limitations, one it is tremendously sensitive to noise and other one is that it increases high
contrast areas much more than area that do not show high image dynamics. Therefore output
image suffers from unkind overshoot objects. Adaptive Unsharp Masking hires an adaptive filter
that controls the contribution of sharpening in the manner that contrast enhancement occurs in
high dense areas and less or no image sharpening occurs in soft areas. This algorithm is better
compared with several other methods available in linear unsharp masking filter technologies.
Binarization is the initial step for processing, with the fact of degradation of the source document,
whichever global or local thresholding approaches are chosen. The Otsu thresholding algorithm

Computer Science & Information Technology (CS & IT) 97
[3] using the histogram shape analysis, which is most widespread global binarization algorithm.
The thresholding of Otsu yields a promising performance when the histogram has twin modal
distribution. The global threshold is designated automatically by a discriminant standard.
Another method utilize image contrast defined as local image minimum and maximum when
compared with the image gradient process, the image contrast derived by the local maximum and
minimum process has a good property and it is more tolerant to the uneven illumination,
document degradation such as smudge [4]. This method is superior when handling document
images with difficult background variation. Finally, the ancient document image is binarized
based on the local thresholds that are derived from the detected high contrast image pixels when
the same is compared with previous method based on image contrast, the method uses the image
contrast to recognize the text stroke boundary and it can be used to produce high accurate
binarization results.
A common technique for scrubbing the degraded documents is modified iterative global
threshold algorithm [5]. A best approach in the separation of object information from foreground
is to compute a global threshold of intensity value based on which two clusters can be diverted. It
is an iteration approach which can handle many degraded conditions. In each iteration the
intermediate tones are shifted towards background there by providing efficient difference
between foreground and background. It is mostly useful for the documents having non-uniform
distribution of noises.
In order to make foreground inscriptions clearly visible from background, Histogram
normalization is used. The image obtained still may suffer from uneven background intensity
variation which in turn reduces the clarity of the foreground. Further processing of image is done
to get an image with better foreground information. As the intensity of the foreground pixels
differs from the intensity of background, this key factor is used to identify the foreground
characters. The main criteria is to find a threshold value which causes the image components to
lie in one of two levels L0, which is below the threshold value and L1, which is above the
threshold value. The pixels which are above the threshold value represent the nodes of a graph.
The nodes form the basis for representing the foreground. The nodes that are neighbors in the
sampling grid are joined by an edge [6].
This technique is used to reduce the number of pixels in the image by a factor of 4 or 8, which in
turn will decrease the number of pixels that has to be processed. The image is represented as a
graph by associating each pixel to a vertex of a graph and connecting the pixels that are neighbors
in the sampling grid by an edge. The gray value of the pixel is considered as an attribute of the
vertex. Since the image is of finite size so also the graph. Pixels represent the finite regions and
vertices represent the faces. The dual of this graph represents borders of the faces which are inter-
pixel edges and vertices. Dual graph pyramids are constructed by adopting bottom-up approach.
Each level of pyramid represents an adjacency graph where vertices correspond to regions and
edges represent the relation between the regions [7].
The procedure of segmentation has huge importance in the handwritten script identification. Thus
an abundant study of research outcome in related segmentation field was surveyed. The algorithm
based on connected components [8], segmenting the document image into non-overlapping equi-
width vertical zones.

A method is to decrease the noise level, which is existing in the distorted image is proposed in
[9]. The distorted image will be first binarized using the threshold, which is determined by the
Otsu’s method. Next for each pixel p(x, y) estimate the horizontal and vertical run length count.
If horizontal and vertical run length count is less than a specified threshold then it is assumed to
be noise and will be eliminated.
Gaussian kernel is a linear operation; convolution is used to find the common area between the
profile and the Gaussian kernel. The degree of shift in the Gaussian kernel during the convolution
process linearly varies with horizontal profile information. So this can be used to represent
randomness in the profile and provides a zero crossing smooth curve, when it is convolved with
the profile, represented by ‘C’. The peaks which are above zero are treated as the gaps between
the lines. Based on this information, the line segmentation is performed [10].
Another method for segmentation is nearest neighbor algorithm [11] which is iterative in nature
scans the character from the top left portion of the image. When it reaches a first black pixel, then
the first symbol is identified through the connected component. If it is found to be the first
character of the script, then it will be placed as the first character of the new line used for placing
the character segmentation. The centroid of the character is calculated and stored in an array
separation the x and y coordinators. The document is again scanned from left top to locate the
next black pixel and hence the next character in the document. The centroid of the character also
computed. The distance between the centroid is computed using the distance formula. If the
distance is less than or equal to threshold value then the character is assume to lie on same line.
Otherwise the character is consider being the part of the next line and transferred to the next line
in the result part. This process is continued until all the characters are scanned and whole image
has been traversed. At the end of the iterative algorithm the separated lines are obtained from the
source. The individual character can also be obtained in this process itself.
Text line segmentation is necessary to detect all text regions in the document image. The
algorithm based on multiple histogram projections using morphological operators to extract
features of the image. Horizontal projection is performed on the text image, and then line
segments are identified by the peaks in the horizontal projection. Threshold applied to divide the
text image into segments. False lines are eliminated using other threshold. Vertical histogram
projections are used for the line segments and decomposed into words using threshold and further
decomposed to characters. This kind of approach provides best performance based on the
experimental results such as Detection rate DR (98%) and Recognition Accuracy RA (98%)
[12,13].
Contour tracing is a technique applied to digital images for extracting the boundary of any object.
This kind of system applies one of the recent contour tracing algorithms to separate character by
using the Theo Pavlidis’s algorithm [15]. It works with 4-connected patterns. The width of the
segmented components from this process is checked. If it is more than the criteria value of the
average width of the components, it will be processed in the next stage. This means there are
some touching characters that are not separated [14].
Tracing of Background Skeleton approach, is applied to separate some touching components to
segment touching characters, background skeleton is processed by using the Zhang-Suen thinning
algorithm. Then, contour tracing algorithm is applied to abstract the skeleton of the background.
Subsequently, the characters in each line will be sorted by checking the column position in order

to determine the sequence of the characters. This is applied to practical data from ancient
documents [15, 16].
Enhanced stroke filter have shifted the attention to the skeletons of potential strokes. In this
manner, the task of distinguishing strokes from a complicated background is converted to the task
of comparing the difference between skeletons of potential strokes and those of disturbing
patches, both of which can be extracted from the resulting images of previous Stroke filters.
Skeleton constraints such as length and width constraints can be introduced into the method to
enhance stroke information [17].
For the segmentation of unconstrained handwritten connected numerals, Water Reservoir
technique is used. A reservoir location and size, touching position (top, middle or bottom) is
decided. Analyzing the reservoir boundary, touching position and topological structures of the
touching pattern, the best cutting point and then the cutting path for segmentation is generated
[18].
Segmented linked characters are critical preprocessing steps in character recognition applications.
Old drop fall algorithm has proved to be an efficient segmenting method due to its simplicity and
effectiveness. However it is subject to small convexes on the contour of characters. Xiujuan
Wang, Kangfeng Zheng, and Jun Guo presented Innertial Drop fall algorithm and big drop fall
algorithm to avoid this defect [19, 20, 21].
3. SYSTEM ARCHITECTURE
The system “Enhancement and Segmentation of Historical Records” designed, mainly consists of
the subcomponents - Preprocessing and Segmentation as shown in Figure 1. The input to the
system is ancient epigraphic documents of varying amount of degradation.
• Image Enhancement : This sub system enhances the quality of the ancient document
images by reducing noise. This is carried out by providing four different filtering options
of various filter sizes.
• Binarization: This sub-system converts RGB images to binary images using
thresholding technique known as Otsu algorithm.
• Segmentation: This sub-system samples out characters from the ancient documents and
is achieved through Drop Fall and Water reservoir techniques.

Figure 1. Enhancement and Segmentation System
4. REVIEW ON THE APPROACHES USED IN PROPOSED SYSTEM
The methods used in current system are described in this section:
Preprocessing stage of degraded ancient document images includes: enhancement and reduction
in noise, which is achieved through different filtering methods for smoothing or sharpening
namely Bilateral, Mean, Median, and Gaussian Blur Filters. These filters are provided with
different mask sizes and parameter values. This is followed by binarization of the enhanced
image to highlight the foreground information, using Otsu thresholding algorithm. Character
segmentation is performed based on Drop Fall and Water Reservoir concept.
4.1 Mean Filter
The mean filter is a sliding-window longitudinal filter that exchanges the center value in the
window with the average (mean) of all the pixel values in that window [22]. Let Sxy represent
the set of coordinates in a rectangular sub image window of size m x n, centered at point (x, y).
The arithmetic mean filtering process computes the average value of the corrupted image g(x, y)
in that area defined by Sxy. The restored image value at any point (x, y) is merely the arithmetic
mean calculated using the pixel in that region, indicated in Equation 1.
݂෩ሺ‫,ݔ‬ ‫ݕ‬ሻ =
ଵ
௠೙
∑ ݃ሺ‫,ݏ‬ ‫ݐ‬ሻሺ௦,௧ሻ∈ௌೣ೤
(1)
This operation can be applied using a convolution mask in which all coefficients have value
1/mn. Mean Filters smoothes local variations in an image and as a result of blurring, noise is
reduced.
4.2 Median Filter
In image processing, neighborhood averaging is the best method to perform the noise reduction,
whereas the method can overturn isolated out of range noise, however the adverse effect is that it
also distorts sudden changes such as sharp edges. The median filter can suppress the noise
without damaging the sharp edges. In median filtering, all the pixel values are first sorted into
numerical order and then replaced with the middle pixel value [22].

Let y be a pixel location and w a neighborhood centered on location (m, n) in the image, therefore
median filter is given by Equation 2,
y [m, n]=median{x[ i ,j], ( i , j) belongs to w} (2)
Subsequently the pixel y [m, n] represents the location of the pixel y, m and n represents the x
and y co-ordinates of pixel y. w represents the neighborhood pixels surrounding the pixel position
at (m, n), (i, j) belongs to the same neighborhood centered on (m, n). Hence the median method
will take the median of all the pixels within the range of (i, j) represented by x [i, j].
4.3 Gaussian blur Filter
A Gaussian blur or Gaussian smoothing involves blurring an image by Gaussian function and
used to decrease image noise and image details. Gaussian smoothing is used as a preprocessing
stage in computer vision algorithms in order to enhance image structures at different scales. The
2- dimension Gaussian function is given by Equation 3.
Gሺx, yሻ < −
ଵ
ଶπσమ e ି ୶మା୷మ
ଶσమ (3)
where x is the distance from the origin (Horizontal axis), y is the distance from the origin
(vertical axis), and σ is the standard deviation of the Gaussian distribution. The standard
deviation σ of the Gaussian determines the amount of smoothing [22].
4.4 Bilateral Filter
Bilateral filter [23] is a non linear filter in spatial domain, which does averaging without
smoothing the edges. The bilateral filter inputs a weighted sum of the pixels in a local
neighborhood; the weights depend on both the spatial distance and the intensity distance.
Essentially the bilateral filter has weights as a product of two Gaussian filter weights, one of
which corresponds to average intensity in a spatial domain, and second weight corresponds to the
intensity difference. Hence no smoothing occurs, when one of the weights is close to 0, which
means the product becomes insignificant around the region where intensity changes swiftly,
which represents usually the sharp edges. As a result, the bilateral filter preserves sharp edges
[28]. Pixel location x, bilateral filter output is given in Equation 4
‫ܫ‬ሗ =
ଵ
஼
∑ ݁
ି||௬ି௫||మ
ଶఙ೏
మ − ݁
ି|ூሺ௬ሻିூሺ௫ሻ|మ
ଶఙೝ
మ௬∈ேሺ௫ሻ ‫ܫ‬ሺ‫ݕ‬ሻ (4)
There parameters controlling the fall-off of weights in spatial and intensity domains, respectively.
And are inputs and output images respectively are spatial neighborhood of pixel I(x), and C can
be given as
‫ܥ‬ = ∑ ݁
ି||௬ି௫||మ
ଶఙ೏
మ − ݁
ଶఙೝ
మ௬∈ேሺ௫ሻ (5)
4.5 Otsu’s Method of Binarization
Binarization is the method of converting a grey scale image to a binary image by using threshold
selection procedures to categorize the pixels of an image into either one of the two classes.
Binarization of the image using Otsu method [23] is used to automatically accomplish histogram
shape-based image thresholding or the decrease of a gray level image to a binary image. This
algorithm adopts that the image as thresholded contains two classes of pixels, then calculates the

optimum threshold separating those two classes so that their combined spread (intra-class
variance) is minimal.
In an Otsu's method weighted sum of variances of the two classes is given by:
ߪఠ
ଶ
(t) = ߱ଵ(t) ߪଵ
ଶ
(t) + ߱ଶ(t) ߪଶ
ଶ
(t) (6)
Weights ߱௜ are the probabilities of the two classes separated by a threshold t and ߪ௜
ଶ
variances
of these classes. The class probability is
ߪ௕
ଶ
(t) = ߪଶ
− ߪఠ
ଶ
(t) = ߱ଵ(t) = ߱ଵ(t)߱ଶ(t) [µଵ
(t) - µଶ
ሺ‫ݐ‬ሻ]ଶ
(7)
The class probability ߱ଵሺ‫ݐ‬ሻ is calculated from the histogram as t:
߱ଵ(t) = ∑ ‫݌‬ሺ݅ሻ௧
଴ (8)
While the class mean µଵ
(t) is:
µଵ
(t) = ሾ∑ ‫݌‬ሺ݅ሻईሺ݅ሻ௧
଴ ] ∕ ߱ଵ (9)
where x(i) is ߱ଶሺ‫ݐ‬ሻ the value at the center of the ith histogram bin. Also can calculate ߤଶ on the
right-hand side of the histogram for bins greater than t.
4.6 Drop Fall Algorithm for Segmentation
Drop fall algorithm [24] with respect to the principle that an equally ideal cut between two
touched characters can be created, if one has to role a hypothetical marble off the top of the first
character and create the cut where the marble falls. The important things to be addressed for this
implementation are where to drop the marble from because it is important if the algorithm starts
at the wrong place. The marble can simply roll down the left side of the first digit or the right side
of the second digit and, hence, it would be completely unsuccessful. The best approach to start
drop falling process is possible to the point at which two characters are touched. In this process
the pixels are scanned row by row until a black boundry pixel with adjacent black boundry pixel
to the right of it is identified, where as the two pixels are separated by white space. This pixel is
used as a point to start the drop fall as shown in Figure 2.
Figure 2. Identification of Initial Pixel Positions

The direction that the algorithm will move is according to the current pixel position and its
surroundings as shown in Figure 3.
Figure 3. The Principle of Drop Fall algorithm
4.7 Water Reservoir Algorithm
The larger space generated by touching characters is analyzed with the help of water reservoir
concept. The working principle of water reservoir method is illustrated in Figure 4. When water
is poured from top (bottom) of a component, the regions of the component where water will be
stored are considered as top (bottom) reservoir. Top (bottom) reservoir is the reservoir obtained
when water is poured from top (bottom). The white spaces are found in the regions in the
bounding box of the components where water can be stored. These regions are called water
reservoirs. The reservoirs obtained in this procedure are not considered for further processing.
Those reservoirs whose heights are greater than a threshold value T1 are considered for further
processing.
Figure 4. Reservoirs formed from water flow from top and bottom is shown for (a) top (b) middle and (c)
bottom touching numerals. The top Reservoirs are marked by Dots and bottom reservoirs are marked by
small line segments

When two characters touch each other, they create a space (reservoir) between the characters.
This space is very important for segmentation because,
a) As cutting points are concentrated around the base of the reservoir, and hence, decreases the
search area.
b) The cutting points lie on base of the reservoir.
c) The space attributes (center of gravity and height) aid to go near the best touching position.
If water is poured from top (bottom) of large space created by touching (Water reservoir) Base of
the reservoir connected numeral then water will be stored in this large space. This water stored
area is named “Water Reservoir” [25]. Figure 5 illustrates the same.
Figure 5. Examples of touching numeral and Space created by the touching
Figure 6 depicts the detection of touching position recognition. The largest reservoir of the
component whose center of gravity lies in vm region is found. This reservoir is known as the best
reservoir for touching. The base-line (lowermost row of the reservoir) of the best reservoir is then
identified. The best reservoir and its base-line are shown and to find this touching position in the
components, morphological thinning operation is applied to touching components for further
processing. For feature points extraction the touching position is renowned. The leftmost and
rightmost points of the base-line of considered reservoirs are the feature points. These points are
initial feature points. With this initial feature points the best feature point (which gets maximum
confidence value) is chosen for segmentation. To calculate confidence value (CV) following
features are considered. Euclidean distance of feature points from the center of gravity of the
touching component.
Figure 6. Feature Detection Approach.

5. PROPOSED SYSTEM AND METHODOLOGY
5.1 DFD of the Preprocessing and Segmentation System
The Data Flow Diagram(DFD) of the current system is shown in Figure 7. This work is carried
out in two phases. In the first phase - Preprocessing, the degraded ancient document image is
taken as input and it is converted to grayscale. Then the smoothing or sharpening filters, namely
Median filter, Gaussian filter, Mean filter, Bilateral filter of different mask sizes are applied to
reduce the amount of noise and thus enhances the image. Next, the enhanced image is binarized
using Ostu algorithm to differentiate background and foreground of ancient document.
In the second phase Segmentation is carried out using Drop Fall algorithm and Water Reservoir
algorithm, to obtain sampled characters, which can be used later stages of OCR.
Figure 7. DFD of the Preprocessing and Segmentation Phases
5.2 Methodology
The functionality of the phases – Preprocessing and Segmentation in detail is covered in this
section. The ancient input image is first converted to grayscale image. The image is enhanced and
noise is reduced by applying four different filters with mask size of 2x2 and 4x4. Next, The
enhanced image is converted to binary image. Lastly the characters in the document are sampled
out using Drop Fall and Water Reservoir approaches.

5.2.1 Preprocessing
Image Enhancement
Input: Degraded Gray Scale image
Functionality: Enhances epigraphic image of medium-level degradation using
spatial filters namely Median, Gaussian blur, Mean, Bilateral filter on the input
image.
Output: The enhanced image with reduced noise.
Algorithm for Enhancement using Median filter
[Step 1]: Read the gray image.
[Step 2]: Compute y [m, n] = median{x [i, j], (i, j) belongs to w}
y be a pixel position, w represent a neighborhood centered around location (m, n) in the
image.
[Step 3]: Apply the Median filter designed over the entire input image to obtain
enhanced image.
Algorithm for Enhancement using Gaussian blur filter
[Step 2]: Compute݁ Gሺx, yሻ < −
ଵ
ଶπσమ e – ୶మା୷మ
ଶσమ
x is the distance from the origin (Horizontal axis), y is the distance from the origin
(vertical axis), σ is the standard deviation of the Gaussian distribution.
[Step 3]: Apply the Gaussian filter designed over the entire input image to obtain
enhanced image.
Algorithm for Enhancement using Mean filter
[Step 2]: Compute ݂መሺ‫,ݔ‬ ‫ݕ‬ሻ < −
ଵ
௠೙
∑ ݃ሺ‫,ݏ‬ ‫ݐ‬ሻሺ௦,௧ሻఢௌೣ೤
Sxy represent the set of coordinates in a rectangular subimage window of size m X n,
centered at point (x, y).
[Step 3]: Apply the Mean Filter designed over the entire input image to obtain
enhanced image.
Algorithm for Enhancement using Bilateral filter
[Step 1]: Read gray image.
[Step 2]: Compute ‫ܥ‬ = ∑ ݁
ି||௬ି௫||మ
ଶఙ೏
మ − ݁
ଶఙೝ
మ௬∈ேሺ௫ሻ
[Step 3]: Compute ‫ܫ‬ሗ =
ଵ
஼
∑ ݁
ି||௬ି௫||మ
ଶఙ೏
మ − ݁
ଶఙೝ
మ௬∈ேሺ௫ሻ ‫ܫ‬ሺ‫ݕ‬ሻ
[Step 4]: Apply the Bilateral Filter designed over the entire input image to obtain
enhanced image.
Binarization
Input: Enhanced Image
Functionality: The enhanced images are converted to binary image consisting of
ones and zeroes.

Output: Binarized image
Algorithm for Binarization
[Step 1]: Compute histogram and probabilities of each intensity level
[Step 2]: Initialize initial ߱௜ሺ0ሻ and µ௜
(0)
[Step 3]: Compute through all possible thresholds t=1. Maximum intensity
Revise ߱௜ and µ௜
; Compute ߪ௕
ଶ
(t)
[Step 4]: Desired threshold corresponds to the maximum ߪ௕
ଶ
(t)
[Step 5]: Compute two maxima (and two corresponding thresholds). ߪ௕ଵ
ଶ
(t)
[Step 6]: Compute greater max and ߪ௕ଶ
ଶ
(t) is the greater or equal maximum
[Step 7]: Compute required threshold threshold1+threshold2/2
5.2.2 Segmentation
The segmentation of the document image is carried out at the character level using Drop Fall and
Water Reservoir Approaches.
Input: Binary epigraph image
Functionality: The binarized image is segmented to characters
Output: Segmented characters of the input epigraph.
Drop fall algorithm
Drop falling algorithm forms segmentation path by rolling in between two touching
characters and displays the segmented characters.
[Step 1]: Input the binary image
[Step 2]: Find the Height and Width of the touched characters
[Step 3]: Apply Breadth First Search (BFS) algorithm to find the touched characters
[Step 4]: If found start the Drop fall, the drop falling algorithm it will always move
downwards, crossways down-wards, to the right, or two the left.
[Step 5]: Make the slice where marble parks. Thus Segmentation path for connected
components is found
Water Reservoir Algorithm
[Step 1]: Find the size of the characters to find touched characters.
[Step 2]: The positions and sizes of the reservoirs are analyzed and a reservoir is detected
where touching is made, the initial feature points for segmentation are noted.
[Step 3]: The best feature points are noted from the initial feature points.
[Step 4]: Based on touching position, close loop positions and morphological structure of
touching region the cutting path is produced.
6. EXPERIMENTAL RESULTS, ANALYSIS AND DISCUSSION
6.1 Experimental Results
The system developed is tested on nearly 150 ancient epigraphic images and the results are found
to be satisfactory. The sample experimental results are depicted in following figures. Figure 8
shows the input ancient historical record, which is analysed for different spatial filtering
techniques.

Figure 8. Input Image selected for Pre-processing
Figure 9 shows the results of color to gray scale conversion, when carried out on the image
shown in Fig 8.
Figure 9. The result of Gray Scale Conversion
Figure 10(a) and 10(b) shows the results after Median filtering, for the mask size of 2x2 and 4x4
respectively. The median filter is an effective method that can suppress isolated noise without
blurring sharp edges.
Figure 10(a). The result of Median filtering for Mask size 2x2
Figure 10(b). The result of Median Filter for Mask size 4x4
Figure 11(a) and 11(b) shows the results of Gaussian blur filtering for the mask size of 2x2 and
4x4. The Gaussian blur method is used to blur the sharpen image so that a less edge highlighted
image is produced.
Figure 11(a). The result of Gaussian Blur Filtering for Mask size 2x2

Figure 11(b). The result of Gaussian blur Filtering for Mask size 4x4
Figure 12 shows the results of Mean filtering for the mask size of 4x4. The Mean filter is a
simple filter that replaces the center value in the window with the mean of all the pixel values in
that window.
Figure 12. The result of Mean Filtering for Mask size 4x4
Figure 13 shows the results of bilateral filtering.
Figure 13. The result of Bilateral Filtering
Figure 14 shows the result of binarization, in which the enhanced image is converted to binary
image.
Figure 14. The results of Binarization
Figure 15 and Figure 16 represents the result of Segmentation of Characters using Drop Fall
algorithm and Water Reservoir algorithm respectively.

Figure 15. The result of Segmented Characters using Drop Fall algorithm
Figure 16. The result of Segmented Characters using Water Reservoir algorithm
6.2 Performance Analysis
The dataset includes 150 samples of medium degraded images for preprocessing and
segmentation. The performance of the 2 phases is discussed below:

6.2.1 Filtering Techniques
This system is tested on 150 samples of medium degraded images, using four spatial filtering
techniques of varying mask sizes. The enhancement was found to be appreciable for the mask
size 4x4 for Median filter when the mask size is high then the output image will appear clear with
sharp edges. The mask size of 2x2 for Gaussian blurs results in blurred image. The mask size of
4x4 for Mean filter typically smoothens local variations in an image and noise is reduced as a
result of blurring. Bilateral filter sharpens the edges. The smoothing Gaussian filter will result in
good accuracy if the edge of the input image is very thick, where as in case of sharpening
Bilateral filter gives better output for the medium degraded images.
6.2.2 Segmentation
The system showed good results when tested on 150 varying degraded images, giving
segmentation rate of 85%-90% for Drop Fall algorithm, 85%-90% for Water Reservoir
algorithm.
Figure 17 . Segmentation Rate of Drop fall and Water Reservoir techniques
7. CONCLUSION
The system showed good results when tested on the 150 samples of varying degraded epigraphs.
It provides better enhanced output on applying filters of appropriate mask size - 4x4 mask size
for Median filter, 2x2 mask size for Gaussian blur, 4x4 mask size for Mean and Bilateral filter.
Segmentation is carried out using Drop Fall and Water Reservoir algorithms and system can
efficiently segment characters from ancient document images. System segments the compound
characters correctly when connectivity present. Few cases where in connectivity is absent,
compound character is segmented separately. Segmentation rate of 85%-90% for Drop fall
algorithm and 85%-90% for Water Reservoir algorithm is achieved.

REFERENCES
[1] Tanzila Saba, Ghazali Sulong & Amjad Rehman, “A survey on Methods and Strategies on Touched
Characters Segmentation”, International Journal of Research and Reviews in Computer Science
(IJRRCS), Vol.1, No.2, June 2010.
[2] M. Trentacoste et al. “Unsharp Masking, Counter shading and Halos: Enhancements or Artifacts”,
2012, The Euro graphics Association and Blackwell publishing Ltd. 2012.
[3] Maya R. Gupta, National P. Jacobson, Eric K. Garcia, “Binarization and Image Preprocessing for
searching Historical Documents”, University of Washington, Seattle, Washington 98195, April
2006.
[4] Bolan Su, Shijian Lu, Chew Lim Tan, “Binarization of Historical Document images Using the local
Maximum and Minimum”, pp.9-11, June 2010.
[5] N. Venkata Rao, A.V.Srinivasa Rao, S Balaji and L. Pratap Reddy, “Cleaning of Ancient Document
Images Using Modified Iterative Global Threshold” ,IJCSI international Journal of Computer Science
Issues, Vol.8.Issue 6, No 2, pp. ISSN (online):1694-0814, November 2011.
[6] Dr. B.P.Mallikarjunaswamy, Karunakara K, “Graph Based Approach for Background and
Segmentation of the Image”, Vol 02, pp.ISSN: 2230-8563, June 2011.
[7] Mamatha H.R and Srikanta Murthy K, “Morphological Operations and Projection Profiles based
Segmentation of handwritten Kannada Document”, International Journal of Applied Information
System (IJAIS)-ISSN: 2249-0868 Foundation of computer Science FCS, New York, USA Vol 4-
N0.5, pp. 13-19, October 2012.
[8] Das, Reddy, Govardhan, Saikrishna, ”Segmentation of Overlapping text lines, Characters in printed
Telugu text document images”, International Journal of Engineering science and technology,
Vol.2(11), pp.6606-6610, 2010.
[9] Karthik S, Mamatha H.R, Srikanta Murth y K, “An Approach based on Run Length Count for
denoising the Kannada Characters”, Internatinal journal of Computer applications, Vol 50-No.18,
pp.0975-8887, July 2012.
[10] N. Anupama, Ch. Rupa & Prof. E. Sreenivasa Reddy, “Character Segmentation for Telugu Image
Document using Multiple Histogram Projections”, Global Journal of Computer Science and
Technology Graphics & Vision, Vol 13, 2013.
[11] Srinivasa Rao A.V, segmentation of ancient Telugu Text Documents, published online in MECS ,I.J.
Image, “Graphics and Signal Processing”, pp.8-14 Published Online July 2012.
[12] Partha Pratim Roy, Umapada Pal, Josep Llados, Mathieu Delalandre, ”Multi-oriented and Mutisized
Touching Character Segmentation using Dynamic programming”, 10th international conference on
Document analysis and Recognition, DOI 10.1109/ICDAR.124 IEEE 2009.
[13] Indu Sreedevi, Rishi Pandey, N. Jayanthi, Geetanjali Bhola, Santanu Chaudhury2, “NGFICA Based
Digitization of Historic Inscription Images, ISRN Signal Processing”, Vol 3, 2013.
[14] Pavlidis, T. “Algorithms for Graphics and Image Processing”, Computer Science Press, Rockville,
Maryland, 2009.
[15] Zhang, T. Y. and Suen, C. Y. “A Fast parallel algorithm for thinning digital patterns.
Communication. ACM”, Vol 3, No 27, pp 236-239, 2009.
[16] B Gangamma,Srikanta Murthy K,Arun Vikas Singh, ”Restoration of Degraded Historical Document
Image”, Journal of Emerging Trends in computing and Information Science, Vol 3, No 5, May 2012.
[17] Xiaoqing Lu, Zhi Tang, Yan Liu, Liangcai. “Stroke –based character segmentation of low –quality
images on ancient Chinese tablet”, 12th International conference on document Analysis and
recognition, 2013.
[18] U. Pal, A. Belaid and Ch. Choisy “Touching numeral segmentation using water reservoir concept”,
Pattern Recognition Letters, Vol.24, pp. 262-272, 2003.
[19] S.Khan.“Character Segmentation Heurisrtics for Check Amount Verification “Master Thesis,
Massachusetts Institute of Technology, 1998.
[20] X.Wang, K. Zheng and J.Guo. “Inertial and Big Drop Fall Algorithm”, International Journal of
information Technology, vol.12, No.4, 2006.

[21] Parminder Singh and Harjinder Singh, “ A Comparison of High Pass Spatial Filters using
Measurements and Automation”, Int. J. IJERT, Vol 1, May 2012.
[22] Mamatha H.R, Sonali Madireddi, Srikanta Murthy K, “Performance Analysis of various filters for
De-noising of Handwritten Kannada Documents”, International Journal of Computer Applications
(0975 – 888) Volume 48– No.12, June 2012.
[23] Messaoud, I.B, et.al. “Region Based Local Binarization approach for Handwritten Ancient
Documents, “in Int. Conf. Front. Handwrit recognit. Bari, pp.633-638, 2012.
[24] Srinivasa Rao A V, D R Sandeep, V B Sandeep,S Dhanam Jaya “Segmentation of Touching Hand
written Telugu Characters by using Drop Fall Algorithm, International Journal of Computers &
Technology, Volume 3 No. 2, OCT, 2012.
[25] Praveen kumar. C, Kiran. Y. C, “ Kannada Handwritten Character Segmentation using Water
Reservoir Method” , International Journal of Systems, Algorithms & Application 2012.

Enhancement and Segmentation of Historical Records

More Related Content

What's hot (19)

Similar to Enhancement and Segmentation of Historical Records (20)

Recently uploaded (20)

Enhancement and Segmentation of Historical Records