Selection of an Optimal Set of Features for Bengali Character Recognition

1.1. Background

The necessity to have a workable Bangla Optical Character Recognizer (OCR) is felt by all Bengali speaking people. Bangla (or Bengali), one of the more important Indo-Iranian languages, is the sixth-most popular in the world and spoken by a population that now exceeds 250 million. Geographical Bangla-speaking population percentages are as follows: Bangladesh (over 95%), and the Indian States of Andaman & Nicobar Islands (26%), Assam (28%), Tripura (67%), and West Bengal (85%). The global total includes those who are now in diaspora in Canada, Malawi, Nepal, Pakistan, Saudi Arabia, Singapore, United Arab Emirates, United Kingdom, and United States. The history of developing a complete Bangla Optical Character Recognizer (OCR) dates back to 1990s(Alam & Kashem, 2010; Chaudhuri & Pal, 1998; Mahmud et al., 2003; Omee et al., 2011). Recent trend of digitizing knowledge repositories, implementation of E-Governance requires easy entry of already printed data into computer. A lot of research effort has already been committed to solve many individual intricate issues. A detailed summary is found here (Hossain, Akter, Sarwar & Rahman, 2010). Research toward a commercially viable Basic OCR still deserves considerable amount of effort. A survey of the existing literature exhibits that most of the authors have tried to develop an OCR considering all the steps of a pattern recognition system. For example, all published literature have considered the tasks of preprocessing, noise elimination, skew detection, segmentation, feature selection and extraction, classification and post processing. A complete in-depth focus on any particular aspect, involving the inherent nature of Bangla Language, is still missing in the whole spectrum of research works in this particular area. Here, we have tried to focus on the feature identification part of the whole issue. In our view, this process deserves special attention and care to be paid in order to build a successful OCR. In our knowledge, this is the first attempt so far taken in the domain of Bangla OCR development.

1.2. A Brief on Bangla Alphabet

The basic Bengali character set comprises of 11 vowels, 39 consonants, and 10 numerals. There are also compound characters being combination of consonant with consonant as well as consonant with vowel. A vowel following a consonant sometimes takes a modified shape and is called a vowel modifier. Similarly there are consonants take the shape of a modifier when comes with another consonant. In general, there are 300 basic, modified and compound character shapes. These characters in a word are topologically connected. Bangla is known as an inflectional language. Some of the characters are shown in Box 1.

Basic and modified characters which are about 75 in number occupy about 96% of the text corpus (Chaudhuri, & Pal, 1998). In another way, majority character identification with no or minimal error will ensure more than 90% of efficiency alone. Many characters of Bengali script have a horizontal line at the upper part called ‘matra’ or headline. A Bengali text may be partitioned into three zones shown in Figure 1. The upper zone denotes the portion above the headline, the middle zone covers the portion of basic characters or compound characters below the headline and lower zone is the portion where some of the modifiers can reside.

Figure 1.

Partitioning line into zones

Some common properties in Bengali language are given below (Chaudhuri & Pal, 1998; Das et al., 2010):

• •

  Bengali script flows from left to right.

• •

  Characters are not classified as uppercase or lowercase.

• •

  A vowel with a consonant takes a modified shape known as vowel modifier.

• •

  A consonant with a consonant takes a modified shape known as consonant modifier.

• •

  The vowel (A) always occurs at the beginning of the word.

• •

  There are about 250 compound characters.

• •

  Many characters have a “matra” or headline with them. Some characters have a signature extended above the headline (e.g. B, C, D).

• •

  In a standard text piece, 95.63% percent of the characters are basic characters and the rest 4.27 percent are compound characters.