SIGNATURE WORK
CONFERENCE & EXHIBITION 2024

This project introduces a cutting-edge sign language recognition system aimed at improving digital interactions for the deaf and hard-of-hearing community. At the heart of this endeavor is a deep learning model that combines high-resolution posture estimation with attention mechanisms that focus on both spatial and temporal aspects of sign language. The model excels in accuracy, with training results boasting a high average of 93.36% correctness and a low loss rate. However, when tested, the accuracy dipped to 51.23%, which suggests that while the model is strong, there’s room for growth, especially in generalizing to new data.

One of the key challenges identified is the limited data for certain sign language categories in the WLASL dataset, which affects the model’s learning depth. To combat this, I’m exploring the use of Generative Adversarial Networks to expand our dataset, making it richer and more diverse. I’m also working on refining how the model processes physical postures to ensure no valuable information is lost during translation from movement to meaning.

Ultimately, my goal is to break down the communication barriers that often sideline the deaf community in digital spaces. This project is more than academic to me-it’s a step towards a world where everyone can connect freely, without limitation. With each improvement, I aim to push the boundaries of technology, making it a welcoming space for all.