A Bilingual Numeral OCR System for Creating UniLingual Digitized Numeral Document

The optical character recognition has been used in many applications such as dictionary generation, customer billing system, banking and postal automation, and library automation etc. The bilingual OCR system to make uni-lingual script helps us to reduce the requirement of two different OCR systems into a single OCR system for recognition of two different languages. This type of globalization helps the universal users of any language can read the text documents in their self-language if the bilingual documents are converted into uni-lingual document. In this paper, the image which contains printed Tamil and European numerals has been recognized using common OCR System and the Tamil numerals are converted into European numerals to globalize the document from a bilingual script into a uni-lingual document. The main objective of the work is to bring out the single numeral (European numerals) text document from the input image with two different numerals (Tamil and European Numerals). The Kohonen’s self-organizing map (SOM) based recognition system has been used for recognizing the numerals and recognized characters in bilingual numerals (Tamil and European Numerals) form are converted into Uni-lingual form (European numerals). This paper also discusses the various approaches used for OCR.


Introduction
The development in the field of computers, digital documentation plays a vital role in all filed.The recognition of a bilingual character is a challenging job to the researchers.In this paper, the mixture of isolated Tamil and European numerals image is recognized using Kohonen's self-organizing map and converted into European numerals in order to globalize the bilingual document into uni-lingual document.The digital image of bilingual page has been scanned through optical scanners.
The scanned image is segmented using connected component method after preprocessing.The features of individual glyph of the image are extracted and then the SOM model is trained to recognize the text.The conversion process is followed by the recognition.The Tamil numerals are converted into European numeral.The final output is stored in a text document.The following Figure 1 shows the steps involved in the recognition and conversion process.

Preprocessing
The preprocessing step involves Noise removal and binarization process which is useful for increasing the segmentation performance.The printed bilingual numeral image has been scanned through the optical scanners with the resolution of 100 dpi, and it is stored in the personal computer for further process.The input Bilingual image is shown in Figure 3 which is a scanned image.Initially, the RGB image is converted into a grayscale image.The grayscale image is binarized using Otsu's method (Otsu, N., 1978).In this work, the individual image glyph features which help to achieve a higher recognition rate are measured.Each character attributes like number of horizontal lines, vertical lines, and the number of circles and arcs are defined for all the numerals used in recognition system.The following masks shown in Figure 6 and Figure 7 are used to define the attributes for detecting the horizontal and vertical lines.The horizontal gradient component is calculated by, The vertical gradient component is calculated by, From the above equations, we obtained the gradient strength.The short and long horizontal lines are identified by setting up the threshold values.Similarly the features of vertical lines, arcs and circles are measured.

Recognition
The artificial neural network simplifies the recognition task and is extensively used in the pattern and character recognition (Mumtazimah Mohamad, Md Yazid Mohd Saman and Muhammad Suzuri Hitam, 2013) 2010) discussed about the recognition using SVM, RCS network, Self-organizing Map, Radial Basis Network and Fuzzy Neural Network for their extracted features of Tamil handwritten characters and achieved a higher recognition rate using RCS with the back propagation network.In this paper, Kohonen's self-organizing map (SOM) has been used for recognition task.The SOM is an artificial neural network trained using unsupervised learning method.The SOM uses a neighborhood function for preserving the input space topological properties.To determine the winner unit, the squared Euclidean distance between input and weight vector has been used.

Algorithm
Step 1: Initially the neighborhood parameters and learning rate are set.The weights are initialized.The input vectors are taken based on the attributes defined in feature extraction.
Step 2: For each input vector both Euclidean distance and weights are calculated.
Step 3: The Euclidean distance method is used to determine the winner unit after the input vectors are given.Squared Euclidean distance can be computed using the following expression.
W h e r e X=input vector i=1 to n & j=1 to m.
Step 4: The new weights are calculated as Where β = Learning rate Step 6: Learning rate β is updated.
Step 5: Radius of the topological neighbourhood has been reduced at specified times.
The Figure 8 shows recognized numerals for the given input bilingual image which is shown in Figure 3.The flowchart of the conversion process of the recognized characters is shown in Figure 9. Initially, the total number of characters has been counted.If the recognized character is a Tamil numeral then, the Tamil numeral is converted to equivalent European numeral which is shown in Figure 2. The conversion process is done by simple replacement of existing Tamil numeral.If the recognized numeral is a European numeral then, no replacement takes place.
The conversion process has been done until the character count becomes zero.The uni-lingual numeral output has been shown in Figure 10.Here all the Tamil & European numerals are in European numeral form.

Results
The

Conclusions
The bilingual OCR system to make uni-lingual numeral script has been created.The reduction in the requirement from two different OCR systems for recognizing two different language numerals to the single OCR system has been achieved by our single OCR system.This type of uni-lingual documentation may help for language translation to read the text documents in their self-language.In this paper, the image that contains printed Tamil, and European numerals has been recognized using common OCR System and the Tamil numerals are converted into European numerals.The input of two different numerals is brought into single numeral (European numerals) form.The Kohonen's self-organizing map based recognition system has been used for recognizing the numerals.The various approaches used for OCR also discussed.By using our system, the input of connected numerals can be recognized.In the future, the work can be extended for the rest of the Tamil numerals have distinct representation for 10, 100 and 1000 as ௰, ௱ and ௲ respectively.

Figure 2 .
Figure 2. Equivalent European numerals for the Tamil numerals

Figure
Figure 3. Bilingual Image Input

Figure
Figure 8. Recognized Numerals input image which has printed numerals has been scanned through 'HP Scanjet 200 Flatbed Scanner' with the resolution of 100 dpi.It takes 14-15 Seconds to complete the scanning process.The personal computer of Intel Core 2 Duo CPU E700, 2.8 GHz processor with MATLAB R2013a version has been used to run the program.To train the self-organizing map, the learning rate is assumed as 0.3 and Initial Radius is assumed as zero.The Tamil numerals ௬ & ௯ are unable to recognize because of their similarities of assumed attributes.The various stages of results are shown in each section.

Figure 9 .
Figure 9. Flow chart of Conversion Process Figure 10.Uni-lingual Numeral Output . El-Yacoubi, A., Sabourin, R., Suen, C. Y. and Gilloux, M. (1999) used hidden Markov model-based approach for recognizing the handwritten words.Baum-Welch algorithm has been used for training and Viterbi algorithm logarithmic version carried out for recognition.The unconstrained handwritten characters from single and multiple writer data are recognized using continuous density Hidden Markov models and Statistical Language model by Alessandro Vinciarelli, Samy Bengio and Horst Bunke (2004).Dhandra, B.V., Gururaj Mukarambi and Mallikarjun Hangarge (2012) achieved recognition accuracy of 92.71% using k-Nearest Neighbor algorithm (KNN) and 96.00% using support vector machines (SVM)classifiers for handwritten Kannada vowels.In the same work, they have obtained 97.51% as recognition accuracy for KNN and 98.26% for SVM classifiers for handwritten English uppercase alphabets.Further, they obtained recognition accuracy of 95.77%, and 97.03% is obtained using k-NN and SVM classifiers respectively for mixed characters that contain Kannada vowels and English uppercase alphabets.