A deep learning approach for line-level Amharic Braille image recognition

Area of Science:

• Computer Science
• Artificial Intelligence
• Assistive Technology

Background:

• Braille is crucial for visually impaired individuals' literacy.
• Existing Optical Braille Recognition (OBR) systems face challenges with complex scripts like Amharic.
• Amharic braille characters require two cells, complicating recognition.

Purpose of the Study:

• To develop an effective deep learning model for Amharic Optical Braille Recognition (OBR).
• To address the challenges of half-character identification and character segmentation in Amharic braille.
• To provide a robust OBR solution that minimizes image pre-processing requirements.

Main Methods:

• A deep learning model combining Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) with Connectionist Temporal Classification (CTC) was proposed.
• The model was trained on 1,800 Amharic braille line images and validated on 200.
• Performance was evaluated using Character Error Rate (CER).

Main Results:

• The best-performing model achieved a Character Error Rate (CER) of 7.81% on test data.
• The model utilized 48x256 image dimensions for optimal results.
• The sequence-to-sequence learning approach proved effective without extensive image pre- or post-processing.

Conclusions:

• The proposed deep learning model offers a viable solution for Amharic Optical Braille Recognition (OBR).
• This method simplifies the OBR process, making it more accessible.
• The first Amharic braille line-image dataset was released to the research community.