VML-HP: Hebrew Paleography Dataset

This paper presents a public dataset, VML-HP, for Hebrew paleography analysis. The VML-HP dataset consists of 537 document page images with labels of 15 script sub-types. Ground truth is manually created by a Hebrew paleographer at a page level. In addition, we propose a patch generation tool for extracting patches that contain an approximately equal number of text lines no matter the variety of font sizes. The VML-HP dataset contains a train set and two test sets. The first is a typical test set, and the second is a blind test set for evaluating algorithms in a more challenging setting. We have evaluated several deep learning classifiers on both of the test sets. The results show that convolutional networks can classify Hebrew script sub-types on a typical test set with accuracy much higher than the accuracy on the blind test.