Recognition and Verification of Touching Handwritten Numerals

Jie Zhou IBM Toronto Laboratory 330 University Avenue, Toronto Canada M5G 1R7 E-mail: jiez a.ibm. om Adam Krzyzak, Ching Y. Suen Centre for Pattern Re ognition and Ma hine Intelligen e Con ordia University, Montreal, Canada H3G 1M8 E-mail: fkrzyzak,sueng enparmi. on ordia. a In the eld of nan ial do ument pro essing, re ognition of tou hing handwritten numerals has been limited by la k of good ben hmarking databases and low reliability of algorithms. This paper addresses the e orts toward solving the two problems. Two databases IRIS-Bell'98 and TNIST are built/organized to serve as standard data sets. Working with the samples from these databases, we proposed a Re ognition & Veri ation system measured by pre ision rate, whi h re e ts the system reliability in a lass-spe i manner. The graph-based re ognizer ombines the segmentation-based and segmentation-free approa hes, while the veri er in orporates both general and domain spe i veri ation s hemes. Results supported the e e tiveness of the proposed veri ation s heme. 1 Introdu tion In today's information so iety, businesses and organizations must pro ess immense numbers of business and nan ial do uments that in lude handwritten numerals, su h as bank heques, payment slips, and tax forms. Re ent advan es in re ognition te hnology have allowed the development of automati pro essing systems1;2. However, many problems are still hindering the emergen e of ompletely su essful systems. A parti ular problem is the re ognition of tou hing numerals, most ommonly tou hing numerals pairs. This problem has been gaining the attention of resear hers due to its important pra ti al role 3;4;5, but less work has been done here than on the re ognition of isolated numerals. There are two main obsta les to this resear h: 1. There is a la k of ben hmarking data to support resear h e orts. CEDAR CD-ROM database does ontain some images of tou hing numerals, provided as mixture with isolated numerals. However, it ontains only a few hundred entries. These entries have been olle ted from zip odes, and thus do not represent the features of tou hing numerals typi ally en ountered in nan ial do uments. There is a riti al need for databases that 179