📌 In this repository, you'll find scripts and instructions for both training (from scratch) and testing an ensemble of five character recognition models using a cross––folding technique. It also includes five pre––trained character recognition models derived from experiments detailed in the research paper titled 'OCR Processing of Swedish Historical Newspapers Using Deep Hybrid CNN–LSTM Networks,' which can be accessed here.
📌 Our experiments, conducted with the open––source OCR engine Calamari, showcased the superiority of mixed deep CNN––LSTM hybrid models over previous methods for character recognition of Swedish historical newspaper text dated between 1818 and 1848. The models achieved an average character accuracy rate (CAR) of 97.43%, setting a new state––of––the––art standard in the analysis of 19th––century Swedish newspaper text.
The calamari-cross-fold-train command will train an ensemble of 5 models based on a cross–fold of the provided training data. Training duration is adapted by the --early_stopping_nbest=5 parameter.
usage: calamari-cross-fold-train --network=cnn=80:3x3,pool=2x2,cnn=100:3x3,pool=2x2,lstm=200,dropout=0.5,lstm=200,dropout=0.5 --files path_to_training_data/*.png --best_models_dir some_output_dir --early_stopping_nbest=5
Confidence voting to different predictions of the five best models.
usage: calamari-predict --checkpoint path_to_model_0.ckpt ... path_to_model_4.ckpt --files path_to_test_data/*.bin.png --output_dir some_output_dir
Evaluate the predicted sentences produced by the calamari-predict script against the provided ground truth.
usage: calamari-eval --gt path_to_test_data/*.gt.txt --pred path_to_predictions/*.pred.txt
Calamari 2.1.2 (2022/01/18) https://github.com/Calamari-OCR/calamari
The dataset contains text line images (bin.png) and their corresponding ground truth (gt.txt) from selected Swedish newspapers spanning between 1818 and 1848. The dataset has been further divided into training and test subsets.
test dataset: 1671 lines
training dataset: 6742 lines