Developing a Lightweight Real-world Sign Language Recognition Model Using Hand Energy Image

This thesis presents a lightweight approach for isolated Italian Sign Language (LIS) recognition, leveraging Hand Energy Images (HEIs) and custom convolutional neural networks (CNNs) designed for efficiency and real-world adaptability. Several preprocessing strategies, including Gaussian blur, background lightening, and adaptive thresholding, were explored to enhance model robustness across controlled and unconstrained environments. Experiments conducted on a subset of the Italian Sign Language A3LIS-147 dataset and a real-world dataset recorded by the author revealed distinct trends: models trained without preprocessing achieved the highest performance on the controlled A3LIS test set, while models trained with adaptive thresholding and Gaussian blur achieved superior generalization to real-world data. The best real-world model attained a Top-1 accuracy of 48.28% and a Top-3 accuracy of 75.86%, highlighting the effectiveness of preprocessing in improving transferability. Despite limitations related to dataset size, signer diversity, and isolated sign scenarios, the proposed system demonstrated strong potential for future applications in real-time, mobile, or embedded environments.