Advancing Eye-Gaze Writing through the Integration of Computer Vision and Predictive Text

This study presents a comparative evaluation of two eye-gaze-based text prediction models—YOLOv8 and Haar—both with and without integrating a text suggestion mechanism. The analysis focuses on key performance indicators, including precision, recall, F1-score, and inference time, to assess model accuracy and system efficiency. In the absence of text suggestion, the Haar model demonstrated superior performance, achieving precision, recall, and F1-scores of 0.85, 0.83, and 0.85, respectively, along with an inference time of 52 ms. Conversely, YOLOv8 yielded slightly lower precision (0.88), recall (0.80), and F1-score (0.83), with a higher inference time of ms. However, the integration of text suggestion revealed the advantages of YOLOv8 in terms of scalability and computational stability. Its compact architecture allowed it to accommodate the added complexity without compromising system performance, whereas the larger Haar model experienced frequent crashes. These results indicate that while Haar offers higher accuracy in isolation, YOLOv8 is better suited for real-time applications when text suggestion is incorporated due to its efficiency and robustness. The findings emphasize the critical role of model size and inference speed in embedded systems, particularly in practical domains such as healthcare and assistive technologies. Future research will aim to enhance the YOLOv8 model by leveraging the latest developments in the YOLO framework, integrating online learning for dynamic adaptability and exploring advanced text recommendation techniques to further improve user experience and system performance.