Attention-Based Fully Gated CNN-BGRU for Russian Handwritten Text.

Area of Science:

• Computer Science
• Artificial Intelligence
• Natural Language Processing

Background:

• Handwritten text recognition (HTR) is crucial for digitizing historical documents and improving accessibility.
• Existing HTR models often lack robust performance for low-resource languages like Kazakh and Russian.
• Attention-based encoder-decoder networks have shown promise in sequence-to-sequence tasks.

Purpose of the Study:

• To develop and evaluate a novel deep neural network for handwritten text recognition in Kazakh and Russian.
• To address the limitations of current HTR models for under-resourced languages.
• To establish a new benchmark for HTR performance in Kazakh and Russian.

Main Methods:

• Development of a novel deep neural network architecture: Attention-Gated-CNN-BGRU.
• Utilizing fully gated Convolutional Neural Networks (CNN), Bidirectional Gated Recurrent Units (BGRU), and attention mechanisms.
• Training and evaluation on handwritten text databases including English, Russian, and Kazakh datasets.

Main Results:

• Achieved a Character Error Rate (CER) of 0.045 and Word Error Rate (WER) of 0.192 on the first test dataset.
• Achieved a CER of 0.064 and WER of 0.24 on the second test dataset.
• Demonstrated statistically significant improvements (p-value < 0.05) in sensitivity (recall) compared to other models on the Handwritten Kazakh and Russian (HKR) dataset.

Conclusions:

• The proposed Attention-Gated-CNN-BGRU model is effective for handwritten text recognition in Kazakh and Russian.
• This work represents the first successful application of deep learning for HTR in these languages.
• The model offers superior performance on the HKR dataset, outperforming existing well-known models.