Abstract
Communication is a critical skill for humans. People who have been deprived from communicating through words like the rest of humans, usually use sign language. For sign language, the main signs features are the handshape, the location, the movement, the orientation and the non-manual component. The vast spread of mobile phones presents an opportunity for hearing-disabled people to engage more into their communities. Designing and implementing a novel Arabic Sign Language (ArSL) recognition system would significantly affect their quality of life. Deep learning models are usually heavy for mobile phones. The more layers a neural network has, the heavier it is. However, typical deep neural network necessitates a large number of layers to attain adequate classification performance. This project aims at addressing the Arabic Sign Language recognition problem and ensuring a trade-off between optimizing the classification performance and scaling down the architecture of the deep network to reduce the computational cost. Specifically, we adapted Efficient Network (EfficientNet) models and generated lightweight deep learning models to classify Arabic Sign Language gestures. Furthermore, a real dataset collected by many different signers to perform hand gestures for thirty different Arabic alphabets. Then, an appropriate performance metrics used in order to assess the classification outcomes obtained by the proposed lightweight models. Besides, preprocessing and data augmentation techniques were investigated to enhance the models generalization. The best results were obtained using the EfficientNet-Lite 0 architecture and the Label smooth as loss function. Our model achieved 94% and proved to be effective against background variations.