OCR for the Varios corpus

This folder contains the documents we used to create our training models, as well as the training models themselves. The directory is divided into two sub-directories : first, the groundtruths, second, the models developed.

Groundtruths

The ground truth of the Varios corpus was composed with the ground truth of the Moreno corpus to which we added 5 pliegos from the Varios corpus (32 pages), in order to train our model on data in Catalan and with long s. Therefore, the entire GroundTruth data was OCRed once with ABBYY FineReader and then manually corrected. Finally, all the data was exported in PAGE-XML from Transkribus for migration to eScriptorium.

From eScriptorium, the entire Grountruth has been segmented according to the SegmOnto vocabulary, and some baselines have been manually corrected.

PAGE ot Alto

To make our data interoperable, we decided to follow the pipeline developed by the SegmOnto project (segmentation controlled vocabulary, TEI schema, quality control schema). As this project is currently principally focusing on ALTO XML documents, we decided to convert our data into this format from the e-scriptorium platform. We have kept both formats in our repository.

Corpus segmentation

The SegmOnto zones used are :

• MainZone, in pink.
• MainZone:columns, in orange.
• RunningTitleZone.
• GraphicZone:illustration, in dark green.
• GraphicZone:ornamentation.
• MusicZone (some of our chapbooks contain sheet music).
• QuireMarksZone:signature.
• QuireMarksZone:catchwords.
• NumberingZone:page.
• MarginTextZone:colophon in light green, this zone isolates the name of the printer, the place and date of printing.
• MarginTextZone:note.

To match with the specificities of our corpus, we have used CustomZone with our own subtypes, corresponding to our choices of XML-TEI encoding scheme:

• CustomZone:title, in purple.
• CustomZone:impresorNum, in brown, for printer's numbers.
• CustomZone:numer_pliego, when there is mention of the number of pliegos assembled.

For the lines we use : - CustomLine with the subtype :trailer to indicate explicit. - HeadingLine. - DefaultLine.

Split

We chose to manually divide our Groundtruth into three sets (80% train, 10% eval and 10% test) in order to ensure that the title pages (where the error rate is often higher due to typography) and the new elements of the Varios corpus are distributed equally. Each set was made up of about 20% of title pages and chapbooks belonging to the Varios corpus. Predefining these three sets also ensures that the results of our different models can be compared.

Groundtruth test

A series of tests was done to evaluate the interest of image processing (binarisation, deblurring, luminosity) for character recognition. One can find more information here : Readme Groundtruth test.

The results obtained imply that post-processing with Niblack binarisation gives the best results on the OCRisation of our corpus ( 96.80 %). Regarding segmentation, the Niblack binarisation does not bring better results.

Models

Corpus Out of domain

In order to improve our results, we undertook a fine-tuning of our model. The corpus used was constituted from the data of the HTRCatalogs established by Juliette Janes, Simon Gabay and Batrice Joyeux-Prunel, whose segmentation was corrected to correspond to the latest standards of SegmOnto.

Modification of the Architecture: The BEST HTR Model

The BEST model yielded the most accurate results. It was trained on an expanded dataset, which included the original Ground Truth enriched with 250 corrected Varios documents. The script used, submission-script-architecture_+.sh, modified Krakens architecture by adding multiple convolutional and LSTM layers while setting low learning rate parameters.

Transcription model accuracy: 96.7%

The Best Segmentation Model

For segmentation, better results were achieved using the YALTAi approach developed by Thibault Clrisse.

Segmentation model accuracy: 86%

Contact

For more information, feel free to contact us at demelerlecordel[at]gmail.com.