Intermediate results of experiments dealing with the training of mixed models for historical printed and handwritten material which seem to be somewhat promising/useful.

The deep3 prefix means that the model was trained using a considerably deeper network structure compared to the default one. Cf. the upcoming selection of default networks for details.

Most of the models focus on printings using Latin script and Antiqua/Fraktur types. Training data included works from the 15th to 19th century. Please see the corresponding paper for details. The basic model is LSH-4 (Latin Script Historical) which was also used as a starting point to refine the other Antiqua and Fraktur models.

The "htr" models focus on the recognition of medieval German manuscripts in gothic and bastard fonts. More details about the training data etc. to follow.