Text Line Segmentation in Images of Handwritten Historical Documents

Abstract

This paper describes an original method to segment handwritten text lines from historical document images. After an initial preprocessing, we compute a black/white transition map to achieve a rough detection of the line regions in the image. Using this map, the corresponding line axes are extracted through a skeletonization algorithm and the conflicts between adjacent cutting lines are solved by some heuristics. Our approach was tested on a set of handwritten digitized documents (from the PROHIST Project database) from the end of the 19th century onwards. The proposed method worked well even with difficult images and it achieved an 82.18% of correct segmented lines for our database. The results of comparing our method with other recent proposal for automatic line extraction on the same test images offered more than a 38% of correct segmentation improvement.

Text Line Segmentation With Water Flow Algorithm Based on Power Function

“…Hence, it is performed over different reference samples of text closely related to handwritten text elements as well as the real ones. Experiments framework for the evaluation of algorithm's text line segmentation consists of the following tests [34,35] [36].…”

“…• , 10 34,35]. Furthermore, between line spacing is set to standard values to 20 % of the standard character height [37].…”

Text Line Segmentation with the Algorithm Based on the Oriented Anisotropic Gaussian Kernel

“…After the algorithms application, the number of connected components is changed. [16]. On contrary, undersegmentation text lines incorporate the mutually joined lines.…”

“…On contrary, undersegmentation text lines incorporate the mutually joined lines. In this case, the sequence of n consecutive text lines is considered as a unique line [16]. It is counted as 1 correctly segmented text line and n − 1 under-segmented text lines [16].…”

“…In this case, the sequence of n consecutive text lines is considered as a unique line [16]. It is counted as 1 correctly segmented text line and n − 1 under-segmented text lines [16]. Mixed-segmentation text lines are linked with incorrectly assigned words between two adjacent text lines.…”