kraken-ocr-data

🔥 @jeandamien-genero — 🚒 special thanks to @Lucaterre

About

This repository is made to present data used or made up in an ocr attempt using Kraken ocr system. The initial purpose was to transcript 3 chapters of Les ouvriers des deux mondes (Volume 3, book 3) published in 1913. It is a collection of french sociology surveys of the early XXth century usually referred to Le Play's monographs (french les monographies de Le Play).

The original book can be found in the Princeton University Library. It was digitized by Google Books and, thanks to the HathiTrust Digital Library, is accessible at there.

The work was conducted by Jean-Damien Généro, engineer at Centre Maurice Halbwachs (affiliate to the École normale supérieure and CNRS), for the research program "TIME US".

Content

• ./scripts/ : scripts used in the process (one bash script and three python functions) ;

• ./training_data_sample/: only one page from the training data (.jpeg, .tiff), the ground truth (.txt) and the segmentation files made up by Kraken out of output_109a.html data (.txt & .png).

• output_109a.html : file resulting of the ketos transcribe command (segmented images and transcription).

• terminal_kraken_training.txt : copy of terminal data during the training.

• model_best.mlmodel : best model resulting of Kraken training.

Data

• 80 pages, 3150 segments and ground truth transcriptions.

• 14 epochs, best model is 98% accuracy report.

Process

Step one : getting images and binarzing them

Images have been downloaded from the HathiTrust Digital Library and binarized using kraken (see binarize function).

Step two : getting ground truth transcriptions

Ground truth transcriptions are needed to perform training. For this purpose, many tools could have been used. I chose Transkribus to segment and to automatically transcribe 80 images from monographs 109, 109 bis and 110 ; I then corrected this first transcription by hand.

Transkribus allows exports in ALTO and text (000_ground_truth.txt).

Step three : segmenting with Kraken and implementing ground truth

In a directory containing all .tiffimages, I ran the ketos transcribe -o output.html *.tiff command. It initialized an output.html file containing segmented images and boxes for transcription of each segment (see below picture). I filled them out with ground truth from Transkribus with the help of a Python Beautifull Soup script (training_data).

Step four : getting training data and perform actual training

I ran the ketos extract --output output_directory *.htmlcommand, which analyzes the output.htmlfile and creates a pair of .png and .txt files containing image of a segment and its ground truth transcription in a new directory (output_directory).

After this, I performed the actual training by running the ketos train *.png command in the output_directory. Fourteen epochs have been necessary to complete the training. For each epoch, a model was created (.mlmodel) ; Kraken only stops when the error rate stops increasing significantly. It then choose the best model, which was epoch 9 (98% accuracy report). You can check out all the process in the terminal_kraken_training.txt file.

As I didn’t have access to a computer cluster as I was used to at Inria Paris, this step was very tideous and took me up to 3 hours.. Make sure you don't have anything important to do on your computer before running ketos train !

Final step : using the best model to perform transcriptions

I am now able to transcribe an image with kraken, by running this command : kraken -i [img file] [output file] segment ocr -m model_best.mlmodel.

For an unknown reason, @Lucaterre and I could not find the way to transcribe more than one image with a single command : at each time, the output file was rewritten. We solved this problem by writting a bash script which iterated over every images in a directory batch_recog_kraken.sh.