LeafOCR-Line

Line segmentation is a crucial step in the manuscript digitization pipeline, as segmented lines are directly used for text recognition. A typical technology-assisted workflow involves manuscript collection, image preprocessing for noise removal and contrast enhancement, followed by segmentation of textual objects such as lines, words, or characters, and finally recognition using machine learning or deep learning methods. However, historical palm leaf manuscripts exhibit varying levels of deterioration—minimal, moderate, and severe—making character-level segmentation impractical, especially for moderately to highly deteriorated documents. Therefore, line-wise recognition is more effective and reliable for such manuscripts. This, in turn, requires accurate line segmentation and well-annotated datasets, which are currently limited in both size and diversity. LeafOCR-Line is a dataset designed to improve line segmentation for palm leaf manuscript digitization and OCR. It contains 1,710 Malayalam palm leaf images collected from 20 literary works, along with corresponding ground-truth masks. Images and masks follow the naming format <i>manuscriptname_pagenumber</i> with .jpg and .png extensions, respectively, and are organized into train, validation, and test folders under separate image and mask directories. The dataset includes 405 highly deteriorated, 911 moderately deteriorated, and 394 less deteriorated images. The training set has 1,198 images, validation 255, and test 257, distributed proportionally across deterioration levels.