Extracting local reliable text regions to segment complex handwritten textlines

Abstract

Textline segmentation is an important preprocess before trying to recognize words. Handwritten texts include complex lines such as connected/overlapped, multi skewed, and curved textlines. In the proposed approach, to overcome these problems, local reliable text regions are locally extracted for each block of a handwritten text. Text image is first filtered by a set of directional 2D filters and filtered images are divided to a number of overlapping blocks. The filtered block with the highest contrast is selected to be used for text region detection. Experiments show that our proposed method accurately segments complex handwritten textlines.