Thresholding, Noise Reduction and Skew correction of Sinhala Handwritten Words

The Sinhala script, which is generally with round characters, is unique among other Brahmi-descended scripts and is used by 70% of the 18 million populations in Sri Lanka. There has been no published research on the cursive unconstrained Sinhala handwriting recognition. This paper proposes vital preprocessing stages, which are categorized under thresholding, noise removal, skew detection and correction algorithm and is useful to improve accuracy of segmentation and recognition of the Sinhala handwritten words. This paper introduces a novel skew detection method based on least square method and also robust indirect skew correction method of unconstrained cursive Sinhala words. Threshold selection is used in combination with three methods such as QIR, NIR, and analyzing gray level intensity distribution. Median filtering algorithm and connected component analysis method are used to reduce the noise in Sinhala handwritten word images and used vertical projection profile histogram based on validation technique to improve the noise reduction of the image. In this proposed system over 700 handwritten patterns in Sinhala handwritten real postal addresses in NSF database [3] were tested and the reported accuracy was 97.2% and 500 handwritten patterns, which were written by undergraduate students, were tested and the reported accuracy was 99.6%. Overall accuracy of 98.4% was reported using these two types of handwritten patterns.