Doc-UFCN - Generic historical line detection

The generic historical line detection model predicts text lines from document images.

Model description

The model has been trained using the Doc-UFCN library on 10 historical document datasets including these public datasets:

It has been trained on images with their largest dimension equal to 768 pixels, keeping the original aspect ratio.

Evaluation results

The model achieves the following results on the test sets:

dataset	IoU	F1	AP@[.5]	AP@[.75]	AP@[.5,.95]
Bozen	60.15	75.10	97.14	3.79	27.50
cBAD2017 (READ) Complex	46.79	60.35	56.01	3.40	16.26
cBAD2017 (READ) Simple	53.97	68.43	57.26	8.45	19.39
cBAD2019	50.77	64.52	35.46	2.88	11.51
DIVA-HisDB	41.54	57.88	63.15	0.00	11.69
Horae	48.93	63.95	57.45	5.20	15.55
ScribbleLens	76.61	86.72	98.02	71.87	58.32

The model has been trained to reduce mergers in predictions (see the paper for more details on training). Therefore, despite slightly low evaluation values, the model correctly detects lines on a wide variety of historical and modern manuscript documents.

How to use?

Please refer to the Doc-UFCN library page to use this model.

Cite us!

@inproceedings{boillet2022,
    author = {Boillet, Mélodie and Kermorvant, Christopher and Paquet, Thierry},
    title = {{Robust Text Line Detection in Historical Documents: Learning and Evaluation Methods}},
    booktitle = {{International Journal on Document Analysis and Recognition (IJDAR)}},
    year = {2022},
    month = Mar,
    pages = {1433-2825},
    doi = {10.1007/s10032-022-00395-7}
}

@inproceedings{doc_ufcn2021,
    author = {Boillet, Mélodie and Kermorvant, Christopher and Paquet, Thierry},
    title = {{Multiple Document Datasets Pre-training Improves Text Line Detection With
              Deep Neural Networks}},
    booktitle = {2020 25th International Conference on Pattern Recognition (ICPR)},
    year = {2021},
    month = Jan,
    pages = {2134-2141},
    doi = {10.1109/ICPR48806.2021.9412447}
}