| Optimizing Sign Language Detection with Deep Learning and Preprocessing: A Comparative Study of CNN-Based Models |
| Paper ID : 1034-ICEEM2025 (R2) |
| Authors |
|
Menna Y. Mohamed1, Mohamed E. Mohamed1, Mohamed N. Saad2, Tamer M. Nassef *1 1Faculty of Computer Science, October Modern University for Sciences and Arts (MSA) 2Biomedical Engineering Department, Faculty of Engineering, Minia University, Egypt |
| Abstract |
| Sign language is a visual-manual language used by deaf and hard-of-hearing communities as a primary means of communication, with distinct grammar and syntax separate from spoken languages. Despite its widespread use, the lack of automated sign language recognition (SLR) systems creates barriers in education, employment, and social inclusion. Developing accurate SLR technology is essential to bridge this communication gap and foster accessibility. This work proposes a comprehensive SLR pipeline to address critical challenges like lighting variations, motion blur, and low contrast in sign language images which degrade model performance and limit real-world usability. Addressing these issues is critical for developing accessible communication tools for the deaf and hard-of-hearing community. This work proposes a comprehensive SLR pipeline that combines image enhancement techniques and deep learning to improve robustness across datasets of varying scales and quality. First, we curated the Hand Gesture recognition dataset (HGR) which is a 5000 images dataset showing the hand signs of 4 classes (delete, stop, send, input), six image preprocessing techniques were applied (greyscale conversion, contrast stretching, gamma correction , Contrast Limited Adaptive Histogram Equalization (CLAHE), deblurring, and sharpening) to mitigate noise and improve feature extraction. Five state-of-the-art models (Convolutional Neural Network (CNN), Visual Geometry Group 19-layer network (VGG19), AlexNet, Residual Network (ResNet), Efficient Network (EfficientNet)) were trained and evaluated, with EfficientNet achieving the highest accuracy (94.8%) offering a reproducible framework for future Sign Language Recognition (SLR) research. |
| Keywords |
| Sign language recognition, image preprocessing, deep learning, dataset scalability, human-computer interaction |
| Status: Accepted |