Ghanaian Sign Language Recognition Using Deep Learning

Abstract

Sign Languages, unlike natural languages, involve the use of continuous gestures, body languages, facial expressions and hand movements to convey meaning and most importantly express a signer's thoughts more effectively. Ghanaian Sign Language is the standard sign language used by the deaf in Ghana with a substantial difference to other sign languages as well as cultural conditions that led to its emergence. In this paper, we proposed and implemented a novel yet deep convolutional neural network to classify and recognize Ghanaian Sign Language and attained an accuracy of 96.0%. Further, we leveraged transfer learning techniques by fine-tuning state-of-the-art network architectures pre-trained on the ImageNet database and improved the accuracy with a reported increase of 3.1%. There was no large publicly Ghanaian Sign Language dataset available, so we created our own dataset for evaluation of the proposed convolutional neural network architecture. Conclusively, we plan of extending the dataset with a view of releasing it in the future, subsequently, allowing researches to apply changes to the dataset using image processing and computer vision tools and techniques they consider can be applicable for their task at hand.