CNN-Based Hindi Numeral String Recognition for Indian Postal Automation

Digits/numerals in the Indian pin-code of handwritten postal documents may touch each other and hence digit string recognition is a very challenging task. In this paper, we propose a digit string recognition system for Indian postal documents written in Hindi. Unlike normal text string, in a string of digits there is no contextual information among the digits as a digit may be followed by an arbitrary digit in a string of digits. Because of this, here we propose a new architecture which is based on CNN (Convolutional Neural Network) and CTC (Connectionist Temporal Classification), without using RNN for Hindi numeral string recognition. Also to connect CNN with CTC, we transform the outputs of CNN to a two-dimension vector to meet the feeding requirement of CTC. Furthermore, we utilize dense blocks to build CNN part to extract efficient image features. Comparative studies with the state-of-the-art methods show that the proposed method outperforms the other existing methods on Hindi numeral string recognition.

I. Introduction

Recognition of handwritten numeral string has been a popular research area for many years because of its various application potentials. Some of its potential application areas are postal automation, bank cheque processing, etc. There are many pieces of works on non-Indian character recognition [4], [5]. Although India is a multi-lingual and multi-script country (in India there are about 25 official languages and 11 different scripts are used to write these languages) but not much work is done towards the recognition of Indian handwritten numerals [2], [3], [12], [13], [17].