Segmentation of Bangla words in scene images

  • Authors:

  • Prakriti Banik;Ujjwal Bhattacharya;Swapan K. Parui

  • Affiliations:

  • Indian Statistical Institute, Kolkata;Indian Statistical Institute, Kolkata;Indian Statistical Institute, Kolkata

  • Venue:
  • Proceedings of the Eighth Indian Conference on Computer Vision, Graphics and Image Processing
  • Year:
  • 2012

Quantified Score

Hi-index 0.00

Visualization

Abstract

Some studies on extraction of Bangla texts from scene images are available in the literature. Also, OCR of printed Bangla texts has been extensively studied. However, the performance of available Bangla OCR on scene texts is not acceptable. In this article, we present our recent study of segmentation of characters or their parts from Bangla texts extracted from scene images. The proposed approach detects the background and text by a combination of two algorithms: unsupervised learning algorithm K-means clustering and Otsu's threshold selection. We propose a criterion to choose an optimal K value for K-means clustering. The text segmentation is based on region growing and extraction of both headline and baseline of such texts. These two lines divide a Bangla word into three horizontal zones. The present algorithm segments characters or their parts in each individual zone. This zone-based segmentation approach helps to reduce the number of symbols to be handled by the classifier in the next stage of the OCR system. Our algorithm can also detect an image having only numerals, avoiding zone detection in that case. Extracted scene texts are often affected by artifacts and our segmentation algorithm can remove them efficiently. Our algorithm has been tested on 2460 Bangla words extracted from 260 scene images.