Word level multi-script identification

Abstract

We report an algorithm to identify the script of each word in a document image. We start with a bi-script scenario which is later extended to tri-script and then to eleven-script scenarios. A database of 20,000 words of different font styles and sizes has been collected and used for each script. Effectiveness of Gabor and discrete cosine transform (DCT) features has been independently evaluated using nearest neighbor, linear discriminant and support vector machines (SVM) classifiers. The combination of Gabor features with nearest neighbor or SVM classifier shows promising results; i.e., over 98% for bi-script and tri-script cases and above 89% for the eleven-script scenario.

Introduction

Demand for tools with capability to recognize, search and retrieve documents from multi-script and multi-lingual environments, has increased many folds in the recent years. Thus, recognition of the script and language play an important part for automated processing and utilization of documents. Plenty of research has been carried out for accomplishing this task of script recognition at a paragraph/block or line level. While the former assumes that a full document page is of the same script, the latter imagines documents to contain text from multiple scripts but changing at the level of the line. Though the latter is a realistic assumption in some cases, most of the practical situations have the script changing with words. In Fig. 1a and b, we show two text images where the script changes at the word level. Fig. 1a shows a bi-script document, where the presence of interspersed English words in a document of Devanagari script is clearly seen. Similarly, using Fig. 1b we show the variation of script at both line and word level and the presence of three scripts in a document, which is a common occurrence in India.

Most of the optical character recognition (OCR) systems are designed using statistical pattern recognition techniques. It is generally observed that these systems generate good output for specific kinds of documents and when the number of classes is reasonable. Including all the various symbols used for writing in the world, together in one reference set as different classes will be prohibitively high. Most of the Indian scripts have 13 vowels (V) and about 35 consonants (C). As reported by the review paper by Pati and Ramakrishnan (2005a), unlike Roman script, in Indian scripts a consonant combines with another consonant or a vowel to generate a completely different symbol. This is demonstrated in Fig. 2 where two different symbols combine to generate a completely new symbol. Fig. 2a presents the combination of two consonants in Odiya script while Fig. 2b presents a sample case in Devanagari. Thus the consonant–vowel (CV) and consonant–consonant–vowel (CCV) combinations, which appear frequently, generate a huge set of graphemes.

In Telugu script alone, Rajasekaran and Deekshatulu (1977), have identified some 2000 symbols which are used regularly. Kannada, with a very similar script and rules, has comparable number of symbols. Devanagari and similar scripts have close to 6000 such combinations each and other scripts have around 300 symbols. Thus, script identification can act as the preliminary level of filter and reduce the complexity of the search for classifying a test pattern. Moreover, for scripts such as Devanagari and Bangla, its identification decides the further course of processing. This includes removal of the shirorekha,¹ the headline, from the word to separate the different symbols forming the word so that each of them can be individually recognized. Thus, identification of the script is one of the necessary challenges for the designer of OCR systems when dealing with multi-script documents.