Newspaper text recognition in Bengali script using support vector machine

Newspapers contain huge amount of important information on current affairs as well as notable past events. Browsing the digital versions of newspaper documents will become much easier if the documents are indexed or transcribed. To enable the automatic transcription, some computer based systems need to be developed for automatic recognition of newspaper text. However, no such recognition system exists for Bengali script, the second most popular Indian script. This article proposes a newspaper text recognition system in Bengali script for the first time in the literature. Initially, each newspaper article is segmented into image and text portions. Then the text document is segmented into various text lines, then each text line into various words and each word into various characters. Various discriminating features have then been extracted from each character using different feature extraction techniques. The feature vector of each character has then been fed to the support vector machine (SVM) classifier to recognize each character of the newspaper document image. The performance of the proposed system has been evaluated on a self-generated dataset and it provides a text recognition accuracy of 97.78%.