Improved feature reduction framework for sign language recognition using autoencoders and adaptive Grey Wolf...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Automatic Sign Language Recognition (ASLR) systems facilitate communication but face challenges with high dimensionality, leading to computational demands.
• The 'curse of dimensionality' in ASLR results from excessive features, causing prolonged training and high computational costs.

Purpose of the Study:

• To propose an integrated machine learning and swarm intelligence technique to address the dimensionality challenge in ASLR.
• To develop a robust and generalizable ASLR system that enhances accuracy and efficiency.

Main Methods:

• Feature extraction using Histogram of Oriented Gradients (HOG).
• Dimensionality reduction via unsupervised autoencoder.
• Feature set refinement using an improved Grey Wolf Optimizer (GWO) algorithm.
• Classification using a handcrafted artificial neural network (AEGWO-Net).

Main Results:

• AEGWO-Net achieved superior performance, improving accuracy by 6% and F1-score by 4% over PCA-IGWO and KPCA-IGWO.
• The system demonstrated high accuracy (98.40%), F1-score (96.59%), MCC (97.14%), and AUC (96.21%) on six diverse datasets.
• AEGWO-Net showed superiority compared to other existing swarm intelligence techniques.

Conclusions:

• The proposed AEGWO-Net effectively overcomes the curse of dimensionality in ASLR.
• The integrated approach offers a robust and generalizable solution for sign language recognition, improving communication accessibility.