Binarization effects on results of text-line segmentation methods applied on historical documents

Results of document scanning are digital images. Image processing is considered as bidimensional signal processing. Such images are the input of document analysis and recognition systems. Information extraction is one objective of document analysis and recognition systems. Textline extraction is a crucial step of such systems because its output is considered as the input of the recognition step. Most segmentation methods take as input binary images, which explains that binarization methods can affect segmentation results. We study in this paper how does the choice of binarization minimally affects the results of text-line segmentation methods? Several evaluation metrics are used for the comparison between segmentation results. The proposed approach is tested using the benchmarking databases IAM (about 556 images) and IAM historical (about 60 images). The results show that binarization affects the detection rate (DR) and the recognition accuracy (RA) metrics for segmentation evaluation.