Bridging Communication Gaps: A Deep Learning Approach for Peruvian Sign Language Recognition

Moyano M.; Lazarte L.; Chong M. · EDUNINE 2026 - 10th IEEE World Engineering Education Conference: The Role of Engineering and Computer Technology as Artificial Intelligence turns into Reality: A Challenge for Professional Education, Proceedings (2026) · DOI: 10.1109/edunine62390.2026.11547056

Resumen

This study presents a deep learning approach for recognizing the Peruvian Sign Language (PSL) alphabet from video sequences, aiming to promote inclusive and accessible sign language learning. The proposed methodology combines a Convolutional Neural Network (CNN) with a Long ShortTerm Memory (LSTM) architecture to capture the spatial and temporal characteristics of static and dynamic gestures. A custom dataset was created with recordings from ten volunteers in both controlled and uncontrolled environments, covering all 28 alphabet letters. The preprocessing pipeline included hand detection, cropping, grayscale conversion, and contrast enhancement through computer vision filters to improve gesture clarity. Results show that the CNN-LSTM model achieves accuracies above 94% in controlled environments and maintains competitive performance above 84% under mixed real world conditions. Additionally, the use of image filtering consistently improves classification performance, particularly for dynamic gestures. Beyond improving automatic recognition, this model can serve as an assistive learning tool to help users practice and learn sign language.

Instituciones: Universidad del Pacífico