Character Segmentation for Cursive Handwritten Text Using Ligature Classification and Transition Feature

Abstract

AbstractSplitting of touching characters in cursive handwritten text is a critical task in segmentation process. A perfect segmentation of character is required to reduce the error rate of recognition. This paper proposes an approach to segment touching/overlapping and shadow characters in the handwritten text using ligature classification. It falls under the category of dissection method, but does not over segment ‘m’, ‘n’ and ‘u’, where the existing methods do. Binarization is the pre processing step for segmentation, which is performed by global or local thresholding. Sauvola’s method of threshold calculation is employed in this approach to binarize the gray scale image. The skew of the image is corrected by MATLAB code. Statistical analysis of ligature is done, in order to classify the inter-letter links and intra-letter links for evaluating the segmentation points. The Possible Segmentation Points (PSP) is generated based on the transition feature, followed by removal of invalid PSP by incorporating ligature extraction. The integration of transition feature in dissection method avoids unnecessary segmentation points without any attempt of classification and consequently reduces computational cost. A benchmark database IAM is used for fair comparison. The paper exhibits many examples with challenging and normal cases. The experimental results show that the proposed method achieves the segmentation rate of 92 %.KeywordsLigature detection and classificationCore detectionStroke height analysisInter-letter linksIntra-letter links