Evaluating the Impact of Video Enhancement Techniques on a Deep Sign Language Recognition 3d Cnn Model (i3d): a Focus on the Wlasl Dataset

Student: Abdullah Oluwatobi Salami (Project, 2025)
Department of Computer Science
PAN-ATLANTIC UNIVERSITY, KM 52 LEKKI-EPE EXPRESSWAY, IBEJU-LEKKI, LAGOS STATE.

Abstract

This study examines how video enhancement methods affects the accuracy of deep learning-based Sign Language Recognition (SLR) systems. In particular, the study assesses the effects of three enhancement techniques, Histogram Equalization, Contrast Limited Adaptive Histogram Equalization (CLAHE), and Real-ESRGAN, on the classification performance of a pretrained Inflated 3D ConvNet (I3D) model. A controlled pipeline was used to develop and implement the system in Google Colab, where each enhancement method produced a dataset in parallel. Time constraints forced early stopping to avoid overfitting, and performance was compared using top-1 classification accuracy. According to the results, CLAHE-enhanced videos produced the best accuracy (66%) followed by RealESRGAN and the original dataset (62%), while Histogram Equalization performed the worst (47%). These findings suggest that adaptive or AIbased video enhancement techniques can improve gesture recognition accuracy in AI-driven SLR systems. The study highlights the importance of preprocessing in real-world deployments and lays the groundwork for further research into scalable and robust SLR solutions that perform well across diverse video conditions.

Keywords
Sign Language Recognition CLAHE RealESRGAN Histogram Equalization I3D CNN WLASL
Filters
Institutions