Optical Character Recognition System for Amharic

14 Pages·2007·0.23 MB·English

Optical Character Recognition of Amharic
Documents

Million Meshesha C. V. Jawahar
Center for Visual Information Technology
International Institute of Information Technology
Hyderabad - 500 032, India
[email protected] [email protected]

Abstract— In Africa around 2,500 languages character recognition (OCR) is being known
are spoken. Some of these languages have their as one of valuable input devices in this
own indigenous scripts. Accordingly, there is a
respect [1], [2]. This is also supplemented
bulk of printed documents available in libraries,
by the emergency of large digital libraries
information centers, museums and offices.
Digitization of these documents enables to and the advancement of information
harness already available information technology.
technologies to local information needs and Digital libraries such as Universal Digital
developments. This paper presents an Optical
Library, Digital Library of India, etc. are
Character Recognition (OCR) system for
established for digitizing paper documents
converting digitized documents in local
languages. An extensive literature survey and make them available to users via the
reveals that this is the first attempt that reports Internet. The advancement of information
the challenges towards the recognition of technology supports faster, more powerful
indigenous African scripts and a possible
processors and high speed output devices
solution for Amharic script. Research in the
(printers and others) to generate more
recognition of African indigenous scripts faces
major challenges due to (i) the use of large information at a faster rate. These days the
number characters in the writing and (ii) availability of relatively inexpensive
existence of large set of visually similar document scanners and optical character
characters. In this paper, we propose a novel
recognition software has made OCR an
feature extraction scheme using principal
attractively priced data entry methodology
component and linear discriminant analysis,
followed by a decision directed acyclic graph [1].
based support vector machine classifier. Optical character recognition technology
Recognition results are presented on real-life was invented in the early 1800s, when it
degraded documents such as books, magazines
was patented as reading aids for the blind.
and newspapers to demonstrate the
In 1870, C. R. Carey patented an image
performance of the recognizer.
transmission system using photocells, and
Index Terms— Optical Character Recognition; in 1890 P.G. Nipkow invented sequential
African Scripts; Feature Extraction; Classification; scanning OCR. However, the practical OCR
Amharic Documents.
technology used for reading characters was

introduced in the early 1950s as a
I. INTRODUCTION
replacement for keypunching system. A

year later, D.H. Shephard developed the
Nowadays, it is becoming increasingly
first commercial OCR for typewritten data.
important to have information available in
The 1980’s saw the emergence of OCR
digital format for increased efficiency in
systems intended for use with personal
data storage and retrieval, and optical
computers [3]. Nowadays, it is common to and Fuchs [9] present handwritten Amharic
find PC-based OCR systems that are bank check recognition. On the contrary,
commercially available. However, most of there are no similar works being found for
these systems are developed to work with other indigenous African scripts.
Latin-based scripts [4]. Therefore, there is a need to exert much
Optical character recognition converts effort to come up with better and workable
scanned images of printed, typewritten or OCR technologies for African scripts in
handwritten documents into computer order to satisfy the need for digitized
readable format (such as ASCII, Unicode, information processing in local languages.
etc.) so as to ease on-line data processing.
The potential of OCR for data entry II. AMHARIC SCRIPTS
application is obvious: it offers a faster,
more automated, and presumably less In Africa more than 2,500 languages,
expensive alternative to the manual data including regional dialects are spoken.
entry devices, thereby improving the Some are indigenous languages, while
accuracy and speed in transcribing data others are installed by conquerors of the
into the computer system. Consequently, it past. English, French, Portuguese, Spanish
increases efficiency and effectiveness (by and Arabic are official languages of many of
reducing cost, time and labor) in information the African countries. As a result, most
storage and retrieval. African languages with a writing system use
Major applications of OCR include: (i) a modification of the Latin and Arabic
Library and office automation, (ii) Form and scripts. There are also many languages
bank check processing, (iii) Document with their own indigenous scripts and
reader systems for the visually impaired, (iv) writing systems. Some of these scripts
Postal automation, and (v) Database and include Amharic script (Ethiopia), Vai script
corpus development for language modeling, (West Africa), Hieroglyphic script (Egypt),
text-mining and information retrieval [2], [5]. Bassa script (Liberia), Mende script (Sierra
While the use and application of OCR Leone), Nsibidi/Nsibiri script (Nigeria and
systems is well developed for most Cameroon) and Meroitic script (Sudan) [10].
languages in the world that use both Latin Amharic, which belongs to the Semitic
and non-Latin scripts [6], an extensive language, became a dominant language in
literature survey reveals that few Ethiopia back in history. It is the official and
conference papers are available on the working language of Ethiopia and the most
indigenous scripts of African languages. commonly learnt language next to English
Lately some research reports have been throughout the country. Accordingly, there
published on Amharic OCR. Amharic is a bulk of information available in printed
character recognition is discussed in [7] form that needs to be converted into
with more emphasis to designing suitable electronic form for easy searching and
feature extraction scheme for script retrieval as per users’ need. Suffice is to
representation. Recognition using direction mention the huge amount of documents
field tensor as a tool for Amharic character piled high in information centers, libraries,
segmentation is also reported [8]. Worku museums and government and private
offices in the form of correspondence letters, Those African languages using a
magazines, newspapers, pamphlets, books, modified version of Latin and Arabic scripts
etc. Converting these documents into can easily be integrated to the existing Latin
electronic format is a must in order to (i) and Arabic OCR technologies. It is worth to
preserve historical documents, (ii) save mention here some of the related works
storage space, and (iii) enhance retrieval of reported at home and in the Diaspora to
relevant information via the Internet. This preserve African languages digitally
enables to harness existing information (predominantly in languages that use Latin
technologies to local information needs and scripts). Corpora projects in Swahili (East
developments. Africa), open source software for languages
like Zulu, Sepedi and Afrikaans (South
African), indigenous web browser in
Luganda (Uganda), improvisation of
keyboard characters for some African
languages, the presence of African
languages on the Internet, etc. [11].
Therefore, we need to give more
emphasis to those indigenous African
scripts. This is the motivation behind the
present report. To the best of our
knowledge, this is the first work that reports
a robust Amharic character recognizer for
conversion of printed document images of
varying fonts, sizes, styles and quality.
Amharic is written in the unique and
ancient Ethiopic script (inherited from Geez,
a Semitic language), now effectively a
syllabary requiring over 300 glyph shapes
for representation. As shown in Fig. 1,
Amharic script has 33 core characters each
of which occurs in seven orders: one basic
form and six non-basic forms consisting of
a consonant and following vowel. Vowels
are derived from basic characters with
some modification (like by attaching strokes
at the middle or end of the base character,
by elongating one of the leg of the base

character, etc.). Other symbols are also

Fig. 1: Amharic alphabets (FIDEL) with their seven orders available that represent labialization,
row-wise. The 2nd column shows list of basic characters
numerals, and punctuation marks. These
and others are vowels each of which derived from the
bring the total number of characters in the
basic symbol.
script to 310 as shown in Table 1.
representational scheme and classification
Type of Amharic Number of method so as to accommodate the effect of
No. Characters Symbols degradation.
1 Core characters 231
2 Labialized characters 51 B. Printing variations
Printed documents vary in fonts, sizes
3 Numerals 20
and styles. Building character recognition
4 Punctuation marks 8
system is challenging in this situation. For
Total 310
example, some of the commonly used fonts

Table 1. Summary of the number of characters in the in Amharic printed documents include
Amharic script (FIDEL). `PowerGeez', `VisualGeez', `Alphas',

`Agafari', etc. Each of these fonts offers
2.1 Challenges in Building an OCR for
several stylistic variants, such as normal,
African Scripts
bold, italic, etc. They are also written in

different point sizes, including 10, 12, 14,
Character recognition from document
etc. These fonts, styles and sizes produce
images that are printed in African scripts is
texts that greatly vary in their appearances
a challenging task, including Amharic
(i.e. in size, shape, quality, etc.) within
documents. To develop a successful
printed documents. We need to standardize
character recognition system for these
the variation in size by applying
scripts, we need to address some of the
normalization techniques and extract
following issues.
suitable features so that the representation

is invariant to printing variations.
A. Degradation of documents

Document images from printed
C. Large number of characters in the script
documents, such as books, magazines,
The total number of characters in
newspapers, etc. are extremely poor in
Amharic script is more than three hundred.
quality. Popular artifacts in printed
Existence of such a large number of
document images include:
Amharic characters in the writing system is
(a) Excessive dusty noise,
a great challenge in the development of
(b) Large ink-blobs joining disjoint
Amharic character recognizer. Memory and
characters or components,
computational requirements are very
(c) Vertical cuts due to folding of the paper,
intensive. We need to design a mechanism
(d) Cuts at arbitrary direction due to paper
to compress the dimension of character
quality or foreign material,
representation so as to come up with
(e) Degradation of printed text due to the
computationally efficient recognizers.
poor quality of paper and ink,

(f) Floating ink from facing pages etc.
This is the main issue where most
character recognition research even for

Latin and non-Latin scripts also fails. We
Fig. 2: Samples of visually similar characters in Amharic
need to carefully design an appropriate writing system.
E. Language related issues
D. Visual similarity of most characters in African indigenous languages pose
the script many additional challenges. Some of these
There are a number of very similar are: (i) lack of standard representation for
characters in Amharic script that are even the fonts and encoding, (ii) lack of support
difficult for humans to identify them easily from operating systems, browsers and
(examples are presented in Fig. 2). Robust keyboard, and (iii) lack of language
discriminant features needs to be extracted processing routines.
for classification of each of the character These issues add to the complexity of
into their proper category or class. the design and implementation of an optical
character recognition system.

Segmentation Components Feature Feature Classifier
Extraction
Vectors

Refines

Images Model

Preprocessing
Recognizer
Model
Model
Raw
Converted
Images
Text

Document Corrected Post-processing
Images Document Text
Text

Fig. 3: Overview of the Amharic OCR design.

III. RECOGNITION OF AMHARIC SCRIPT information is recorded and, therefore, the
greater the file size and vice versa. The
We built an Amharic OCR on top of an resolution most suitable for character
earlier work for Indian languages [12] and recognition ranges from 200 dpi to 600 dpi
also reported a preliminary result of our depending on document quality. Actually
work in [13]. Fig. 3 shows the general most systems recommend to scan
framework of Amharic OCR design. documents at a resolution of 300 dpi for
The system accepts as input scanned optimum OCR accuracy. We also use this
document images. The resolution, or resolution to digitize documents for the
number of dots per inch (dpi) produced by experiment.
the scanner determines how closely the Once images are captured, they are
reproduced image corresponds to the preprocessed before individual components
original. The higher the resolution, the more
are extracted. Preprocessing is necessary In our implementation, line, word and
for efficient recovery of the text information characters in the text are segmented in the
from scanned image. The preprocessing following manner. Lines are detected from
algorithms employed on the scanned image text blocks using horizontal projection. A
depend on paper quality, resolution of the text line can be found between two
scanned image, the amount of skew in the consecutive boundary lines where the
image, the format and layout of the images height of projection profile is least. Identified
and text and so on [2], [5]. In this work we text lines are then segmented into its
apply binarization, noise removal, skew constituent words.
correction and normalization techniques. We employ vertical projection for word
Scanned documents often contain noise segmentation. This approach scans each
that mainly arises due to printer, scanner, line, the valleys of which show the
paper quality and age of the document. We boundaries of each word in the line. Next,
use binarization to convert gray scale characters are identified using vertical
image (with 256 possible different shades projection. To identify the boundaries of
of gray from black to white pixels) to binary characters each word is scanned in the
colors (with just black and white pixels) to vertical direction. During scanning if there is
ease image processing. In due course no black pixel encountered till the base line
binarization removes some of the noises in is reached, then this scan marks a
images. However, because of the level of demarcation line for characters.
degradations in digitized document images Once character boundaries have been
of magazines, books and newspaper, there detected, it is often useful to extract
is a need to apply noise filters so as to character components. We use connected
reduce the effect of degradation during the component analysis to identify the
recognition process. We use Gaussian connected segments of a character. The
filtering that smoothes an image by final result is a set of character components
calculating weighted averages in a filter box with their bounding box that are scaled to a
[2]. standard size of 20×20 . This dataset is
During the scanning operation, a certain used as an input for feature extraction,
amount of image skew is unavoidable. So, training and testing the recognition engine.
skew detection and correction is
an important pre-processing step. In our IV. FEATURE EXTRACTION
work projection profile have been used for
Feature extraction is the problem of
skew correction in the range of ± 20%.
identifying relevant information from raw
Further detail is available in [12].
data that characterize the component
Once pages are preprocessed, then
images distinctly. There are many popular
they are segmented into individual
methods to extract features. Selection of a
components. Segmentation is an operation
feature extraction method is probably the
that seeks to decompose an image into
single most important factor in achieving
sub-images of individual symbols [14]. The
high recognition performance in character
page segmentation algorithm follows a top-
recognition systems. A survey of feature
down approach by identifying text blocks in
extraction schemes for character
the document pages. This is followed by
recognition is available in [15]. Different
line, word and character segmentation.
feature extraction methods are designed for
different representations of the characters. vector x , where M depends on the image
i
Some of these features consider profiles,
size. From the large sets of training
structural descriptors and transform domain
datasets, x , ,x we compute the
L
representations [2], [16]. Alternates one 1 N
covariance matrix as:
could consider the entire image as the

feature. The former methods are highly
1 N
language specific and become very ∑= ∑(x −µ)(x −µ)T
complex to represent all the characters in N i i
i=1 (1)
the script. The later scheme provides

excellent results for printed character
Then we need to identify minimal dimension,
recognition.
say Ksuch that:
As a result, we extract features from the

entire image by concatenating all the rows
∑K λ/Trace(∑)
to form a single contiguous vector. This i−1 i ≤ α (2)
feature vector consists of zeros (0s) and
ones (1s) representing background and where λ is the ith largest eigenvalue of the
i
foreground pixels in the image, respectively.
covariance matrix ∑ and α is a limiting
With such a representation, memory and
value in percent.
computational requirements are very
Eigenvectors corresponding to the
intensive for languages like Amharic that
largest K eigenvalues are the direction of
have large number of characters in the
greatest variance. The kth eigenvector is
writing.
the direction of greatest variation
Therefore we need to transform the
features to obtain a lower dimensional perpendicular to the first through (k−1)st
representation. There are various methods eigenvectors. The eigenvectors are
employed in pattern recognition literatures arranged as rows in matrix A and this
for reducing the dimension of feature transformation matrix is used to compute
vectors [17]. We consider in the present the new feature vector by projecting as
work Principal Component Analysis (PCA) Y = Ax . With this we get the best one
i i
and Linear Discriminant Analysis (LDA).
dimensional representation of the

component images with reduced feature
4.1 Principal Component Analysis
size.

Principal component analysis (PCA)
Principal component analysis (PCA) can
yields projection directions that maximize
help identify new features (in a lower
the total scatter across all classes. In
dimensional subspace) that are most useful
choosing the projection which maximizes
for representation [17]. This should be done
total scatter, PCA retains not only between-
without losing valuable information.
class scatter that is useful for classification,
Principal components can give superior
but also within-class scatter that is
performance for font-independent OCRs,
unwanted information for classification
easy adaptation across languages, and
purposes. Much of the variation seen
scope for extension to handwritten
among document images is due to printing
documents.
variations and degradations. If PCA is
Consider the ith image sample
applied on such images, the transformation
represented as an M dimensional (column)
matrix will contain principal components ∆ =µ −µ
1 2 (4)
that retain these variations in the projected

feature space. Consequently, the points in
where x is the jth sample in ith class.
the projected space will not be well ij
separated and the classes may be smeared Sum of the with-in class scatter is also
together. determined by,
Thus, while the PCA projections are
optimal for representation in a low A=cW +(1−c)W
1 2 (5)
dimensional space, they may not be optimal

from a discrimination standpoint for
where 0≤c≤1, and the scatter space using:
classification. Hence, in our work, we

further apply linear discriminant analysis to
S =d A−1d
extract useful features from the reduced ij i j (6)
dimensional space using PCA.

4.2. Linear Discriminant Analysis d =αA−1∆
1 1
S−1=1/S
1 11
Linear discriminant analysis (LDA)
for n=2 to K
selects features based entirely upon their d =αA−1{∆−[d d ]S−1[1/α0 0]'}
discriminatory potential. Let the projection n n 1L n−1 n−1 1 L
be y = Dx, where x is the input image and ω =(dt∆)2/dtAd
n n n n

D is the transformation matrix. The 1 ⎡c S−1 +S−1 y ytS−1 −S−1 y ⎤
ombajexcimtivizee so f tLhDe Ar aisti ot oo ffi ntdh ea bperotwjeecetino-nc ltahsast Sn−1=cn ⎢⎢⎣ n n−1−yntnS−1n−−11n n n−1 n1−1 n⎥⎥⎦

scatter and the within-class scatter [15].
We apply the algorithm proposed by Fig. 4: Algorithm for computing L best discriminant
feature vectors. The vertical and horizontal lines are drawn
Foley and Sammon [18] for linear
to partition the scatter matrix S−1 such that the result
discriminant analysis. The algorithm n
extracts a set of optimal discriminant obtained is a 2×2 matrix.
features for a two-class problem which suits
the Support Vector Machine (SVM) The algorithm presented in Fig. 4 is
classifier we used for classification task. called recursively for extracting an optimal
Consider the ith image sample set of discriminant vectors ( dn ) that
represented as an M dimensional (column) corresponds to the first L highest
vector x , where M is the reduced discriminant values such that
i
(ω ≥ω ≥ ≥ω ≥0 ) [18]. Here, K is the
dimension using PCA. From the given sets 1 2 L L
number of iterations for computing
of training components, x , ,x we
L
1 N
discriminant vectors as well as discriminant
compute with-in class scatter (W ) matrix
values, and, α is chosen such that:
and between-class difference (∆) as: n

W =∑Ni (x −µ)(x −µ)t dntdn =1 (7)
i i=1 ij i ij i (3)
y and c are determined in the following classification of unlabelled samples in order
n n
to evaluate classifier’s performance. In
manner.
general it is desirable to have a classifier

with minimal test error.
y =[S S ]
L
n in (n−1)(n) (8) We use Support Vector Machine (SVM)
c =s − ytS−1 y for classification task [19]. SVMs are pair-
n nn n n−1 n
wise discriminating classifiers with the

ability to identify the decision boundary with
Since the principal purpose of
maximal margin. Maximal margin results in
classification is discrimination, the
better generalization [20] which is a highly
discriminant vector subspace offers
desirable property for a classifier to perform
considerable potential as feature extraction
well on a novel data set. Support vector
transformation. Hence, the use of linear
machines are less complex and perform
discriminant features for each pair of
better (lower actual error) with limited
classes in each component classification is
training data.
an attractive solution to design better
Identification of the optimal hyper-plane
classifiers.
for separation involves maximization of an
However, the problem with this
appropriate objective function. The result of
approach is that the classification becomes
the training phase is the identification of a
slower, because LDA extracts features for
each pair-wise class. This requires stack set of labeled support vectors x and a set
i
space for storing N(N −1)/2 transformation of coefficients α. Support vectors are the
i
matrices. For instance, for dataset with 200 samples near the decision boundary. They
classes LDA generates around 20 thousand
have class labels y ranging ±1. The
i
transformation matrices for each pair and is
decision is made from:
very expensive for large class problems.

Hence, we propose a two-stage feature
l
extraction scheme; PCA followed by LDA f(x)=sgn(∑αy K(x ,x))
i i i
for optimal discriminant feature extraction. i=1 (9)
We first reduce the feature dimension using
PCA and then run LDA on the reduced where K is the kernel function, which is
lower dimensional space to extract the most defined by:
discriminant feature vector. This reduces to
a great extent the storage and K(x,y)=φ(x)φ(y)
(10)
computational complexity, while enhancing

the performance of SVM-based decision
where φ:Rd →H maps the data points in
directed acyclic graph (DDAG) classifier.
lower dimensions to a higher dimensional
V. CLASSIFICATION space H.
Binary classifiers like SVM are basically
Training and testing are the two basic designed for two class classification
phases of any pattern classification problem. problems. However, because of the
During training phase, the classifier learns existence of a number of characters in any
the association between samples and their script, optical character recognition problem
labels from labeled samples. The testing is inherently multi-class in nature. The field
phase involves analysis of errors in the of binary classification is mature, and
provides a variety of approaches to solve combine the results of binary classifiers
the problem of multi-class classification [21]. using a suitable voting methods such as
Most of the existing multi-class algorithms majority or weighted majority approaches
address the problem by first dividing it into [21], [22].
smaller sets of a pair of classes and then

Fig. 5: A rooted binary decision directed acyclic graph (DDAG) multi-class classifier for eight-class classification
problem.

5.1 Multi-class SVMs classifiers built for a N class classification
problem is N(N −1)/2. The input vector is
Multi-class SVMs are usually presented at the root node of the DDAG
implemented as combinations of two-class and moves through the DDAG until it
solution using majority voting methods. It reaches the leaf node where the output
has been shown that the integration of (class label) is obtained. At each node a
pairwise classifiers using decision directed decision is made concerning to which class
acyclic graph (DDAG) results in better the input belongs.
performance as compared to other popular For example, at the root node of Fig. 5,
techniques such as decision tree, Bayesian, a decision is made whether the input
etc. [19]. belongs to class 1 or class 8. If it doesnot
We construct a decision directed acyclic belong to class 1, it moves to the left child;
graph (DDAG), where each node in the otherwise, if it doesnot belong to class 8, it
graph corresponds to a two-class classifier moves to the right child. This process is
for a pair of classes. The multi-class repeated at each of the remaining nodes till
classifier built using decision directed it reaches the leaf node where the decision
acyclic graph (DDAG) for 8-class is reached to classify the input as one of the
classification problem is shown in Fig. 5. It eight classes.
can be observed that the number of binary