Categorization of American Sign Language using Convolutional Neural Networks

There are a number of sign languages used around the globe. American Sign Language (ASL) is one of the most widely utilized sign languages. It originated in the early 19th century. The ASL alphabet recognition has the potential to lessen the breakdown in communication among the deaf and the hearing societies. This research includes a convolutional neural network (CNN) model for categorizing ASL letters in images. The model uses a standard CNN architecture and rectified linear unit (ReLU) activation and incorporates multiple convolutional as well as pooling layers to wrest features from the image dataset. We have used an ASL dataset available on Kaggle for public usage which has approximately 87000 images. The dataset contains images for 26 letters from the English alphabet with additional classes including nothing, space, and delete; summing to a total 29 different classes. The proposed model learns on a dataset of ASL alphabet images using Adam optimizer and categorical cross-entropy loss. We have had a very good accuracy result for the model when the performance is assessed on a distinct test set, achieving an accuracy of 96.96%.