Recognition of Arabic Air-Written Letters: Machine Learning, Convolutional Neural Networks, and Optical Character...

Area of Science:

• Human-Computer Interaction
• Artificial Intelligence
• Computer Vision

Background:

• Air writing offers a novel interface for human-computer interaction, with significant research in English and Chinese but limited work in Arabic.
• Bridging the research gap in Arabic air writing is crucial for inclusive technological advancement and metaverse applications.

Purpose of the Study:

• To develop and evaluate a hybrid model for accurate Arabic air writing recognition.
• To enhance human-machine communication through advanced gesture recognition.

Main Methods:

• A hybrid model combining deep convolutional neural networks (CNNs) for feature extraction (VGG16, VGG19, SqueezeNet) with machine learning classifiers (NN, RF, KNN, SVM).
• Integration of optical character recognition (OCR) for segmenting individual characters from continuous gestures.
• Application of grid and random search optimization for parameter tuning on the AHAWP dataset.

Main Results:

• The proposed hybrid model achieved a maximum accuracy of 88.8% using a neural network (NN) with VGG16 features.
• Data preprocessing schemes were employed to enhance data quality and reduce bias.
• OCR integration effectively improved the recognition of isolated Arabic letters.

Conclusions:

• The hybrid deep learning and machine learning approach, enhanced by OCR, demonstrates significant potential for Arabic air writing recognition.
• This research contributes to advancing multilingual human-computer interaction and metaverse technologies.