Attention-based hybrid deep learning model with CSFOA optimization and G-TverskyUNet3+ for Arabic sign language...

Area of Science:

• Computer Science
• Artificial Intelligence
• Natural Language Processing

Background:

• Arabic sign language (ArSL) is vital for communication among Deaf individuals in Arabic-speaking regions.
• Existing ArSL recognition methods face limitations in accuracy and feature extraction.
• Accurate ArSL recognition is crucial for education, healthcare, and societal inclusion.

Purpose of the Study:

• To introduce DeepArabianSignNet, a novel model for enhanced Arabic sign language recognition.
• To overcome the limitations of previous ArSL recognition approaches.
• To improve the accuracy and feature-capturing capabilities in ArSL recognition.

Main Methods:

• Proposed DeepArabianSignNet model combines DenseNet, EfficientNet, and attention-based Deep ResNet.
• Utilized G-TverskyUNet3+ for region of interest detection in ArSL images.
• Employed the Crisscross Seed Forest Optimization Algorithm for feature selection (texture, color, deep learning).

Main Results:

• The model was evaluated on two databases with 70% and 80% training rates.
• Database 2 achieved high accuracy: 0.97675 (70% training) and 0.98376 (80% training).
• Demonstrated significant improvements in Arabic sign language recognition.

Conclusions:

• DeepArabianSignNet proves effective in enhancing Arabic sign language recognition.
• The proposed model addresses previous accuracy and feature extraction challenges.
• This work contributes to advancing assistive technologies for the Deaf community.