Deep Learning-Based Eye-Writing Recognition with Improved Preprocessing and Data Augmentation Techniques.

Area of Science:

• Assistive Technology
• Computer Vision
• Machine Learning

Background:

• Traditional eye-tracking systems (EOG, infrared) are accurate but costly and invasive.
• Vision-based systems are more accessible but under-explored for eye-writing recognition.
• Inconsistent signal lengths from natural eye movements hinder recognition accuracy.

Purpose of the Study:

• To develop a novel, accurate, and accessible vision-based eye-writing recognition system.
• To address challenges of signal length variability and improve recognition robustness.
• To create a new webcam-captured dataset for eye-writing research.

Main Methods:

• Utilized a webcam-captured dataset and introduced Discrete Fourier Transform (DFT) for length normalization.
• Employed a hybrid deep learning model combining 1D Convolutional Neural Networks (CNN) and Temporal Convolutional Networks (TCN).
• Incorporated data augmentation and initial-point normalization for enhanced model robustness.

Main Results:

• Achieved high accuracies: 97.68% on a new webcam dataset, 94.48% on Japanese Katakana, and 98.70% on an EOG dataset.
• The DFT-based normalization standardized input lengths, improving efficiency and robustness.
• The hybrid CNN-TCN model effectively captured spatial and temporal features of eye-writing.

Conclusions:

• The proposed system offers an efficient and robust solution for vision-based eye-writing recognition.
• The novel preprocessing and deep learning approach significantly outperforms existing methods.
• This work advances assistive technology by providing a more accessible communication tool.