Article Preview
Top1. Introduction
In recent years, researchers have become increasingly interested in handwritten document image analysis (DIA). The motivation for that is twofold, first, automating the process of human reading and, second, developing automatic processing systems that can encode the text present in handwritten documents (like postal envelopes, bank cheque, an employee id, etc.) into character codes such as American Standard Code for Information Interchange (ASCII), etc. Despite the omnipresence/proliferation of plastic money, i.e. credit cards, debit cards, smart cards, ach (automated clearing house), and other methods of electronic payment. Paper cheque still has a profound impact/importance, particularly in the non-cash transaction systems (Palacios, 2008). In this system, a bank employee mainly process, the cheques by hand and then authenticates the transaction. So, there is plenty of scope in developing new methods for automating the cheque processing in order to save a considerable amount of time and human effort. The typical sample of a bank cheque is shown in Figure 1. The handwritten fields of it, can be classified as: 1) Legal amount: amount written in words 2) Courtesy amount: the amount written in numeric format 3) Date: current date for issuing a cheque 4) Payee details: the name of the beneficiary, and 5) Signature: gives an authority/permission to a bank to release the money in favor of payee.
As can be observed from Figure 1, the amount in the cheques are deliberately written twice one in words and second in numerals. Any disagreement between these two forms shall be regarded as alteration/fraud. The general view is that the errors made in writing the amount in these two forms are highly uncorrelated (Manuel Leroux, 1997). Thus, paving the way for achieving higher reliability, with respect to amount recognition in cheque processing systems. In other words, a legal amount counter checks the amount written in courtesy format as it is easy to temper with. Therefore, it is an extremely important/pivotal field in cheques and their automatic recognition may speed up the entire domain of automatic bank cheque processing. Over the years, an enormous amount of research has been carried out in an attempt to characterize legal amounts. Like, for example, Latin, Roman/English, Arabic, Persian, French, Chinese, Brazilian, and Japanese etc. (Kolhe & Pal, 2011). Although, different aspects of Indic handwriting has been extensively studied(Pal, Jayadevan, & Sharma, 2012). However, the automatic recognition of legal amounts written in Devanagari script has largely been overlooked. Thus, in this paper, our focus is on recognition of legal amounts written in Devanagari script: an official and popular/major Indic script.
Figure 1. An illustration of various handwritten fields of Bank Cheque (Jayadevan, 2011).
Legal amount recognition, is a special case for the recognition of handwritten words with static/limited lexicons (small vocabularies), only. For example, lexicon sizes in Hindi and Marathi languages of Devanagari script are 106 and 114 words, respectively. Whereas, only 30 to 36 in an English language (Jayadevan, Pal, & Kimura, 2010; Madhvanath & Govindaraju, 2001). In general, handwritten Devanagari word recognition (DWR) can be classified as analytical-based and holistic-based methods. In a case of analytical-based DWR, a handwritten Devanagari word (DW) is segmented into different constituent units such as pseudo-characters, different zones, graphemes, etc., and then trying to recognize each respective individual constituent unit. However, due to the presence of uncertainty in deciding the cut positions, touching characters, variable intra-space between constituent characters in a DW, recognition performance gets negatively affected. Thus, resulting in an over or under segmentation of characters, for example, in cursive handwriting.