Cleaning and Recognition of Numerals in a Handwritten Devnagari Document Consisting of Roll Numbers and Marks

Abstract

The field of character recognition has been fascinating researchers over a last few decades. It is easy for the human mind to decipher the handwritten characters with accuracy, but for the machines, it is a difficult task. The Optical Character Recognition (OCR) systems have been developed to imitate the human action of interpreting the handwritten and printed characters. Innumerable approaches have been developed for the character recognition process but it still remains a challenging field for the researchers. The most of the approaches have used the neural networks as their basic technique for the recognition process. These approaches require huge computations and are training dependent. Thus, an effort has been made by the authors to develop a simple approach for the offline handwritten character recognition. The proposed approach initiates with the preprocessing of the image through various techniques like filtering, binarization, thinning and smoothing. For the purpose of filtering, modified median filter has been proposed, which is based on the conditional application of the median filter. The cleaned up image is then, decomposed into lines, words and characters by using the line segmentation, word segmentation and character segmentation techniques respectively. The segmented characters are analyzed to check the presence of the features like sidebars, closed loops and water reservoirs. Based on these features, a classification structure has been developed, which assigns the segmented character to the predefined class. The prime focus here is on the recognition of the offline handwritten Devnagari numerals, which can aid in several applications like reading commercial forms, postcode recognition, passport readers, bank cheques, bill processing systems and many more. The proposed approach has been implemented. It was applied to several documents and satisfying results have been obtained by the authors.