Modified 3D-Convolutional Neural Network applied in Dynamic Filipino Sign Language Recognition

Advancing communication for the deaf and mute community requires effective sign language recognition systems. This study improves dynamic gesture recognition for Filipino Sign Language (FSL) using a 3D Convolutional Neural Network (3D CNN). Key challenges such as vanishing and exploding gradients, which hinder the model's learning capabilities, were addressed through batch normalization and gradient clipping. Batch normalization stabilized training by reducing gradient variance from 366.8836 to 4.9723. Gradient clipping minimized instability, leading to increased model accuracy. These techniques significantly enhanced model robustness and generalization. This research highlights their importance in developing accessible FSL recognition systems that promote inclusivity and preserve language.