HTR for Greek Historical Handwritten Documents.

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Humanities

Background:

• Historical documents present significant challenges for offline handwritten text recognition (HTR) due to poor manuscript quality and unique historical writing styles.
• Transcribing Greek historical manuscripts is particularly difficult owing to their specific historical particularities.

Purpose of the Study:

• To propose and evaluate a novel deep learning architecture for the accurate offline handwritten text recognition (HTR) of Greek historical manuscripts.
• To address the specific challenges posed by low-quality historical documents and unique historical writing characteristics.

Main Methods:

• A convolutional recurrent neural network (CRNN) architecture incorporating octave convolution and gated recurrent units was developed.
• The proposed CRNN model was evaluated on three newly created collections of Greek historical handwritten documents and standard datasets (IAM, RIMES).

Main Results:

• The proposed architecture demonstrated effective performance in transcribing challenging Greek historical manuscripts.
• Comparative analysis showed the proposed model outperforms existing state-of-the-art architectures on these specific datasets.

Conclusions:

• The developed convolutional recurrent neural network architecture offers a robust solution for the offline handwritten text recognition (HTR) of historical Greek manuscripts.
• The model's effectiveness in handling challenging historical document features paves the way for broader applications in historical text analysis and digitization.