Unlocking Communication: ASL Fingerspelling Recognition using Deep Learning Algorithms

Fingerspelling in American Sign Language (ASL) is an important means of communication for the Deaf and hard-of-hearing community, but limitations in awareness of ASL as a form of communication and other technological factors lead to communication limitations. The project seeks to improve ASL fingerspelling identification through deep learning techniques for easier, more accessible communication. Evaluation is conducted on five advanced convolutional neural network architectures-VGG16, ResNet50, MobileNetV2, DenseNet121, and InceptionV3 with a very large dataset consisting of more than three million hand gesture samples from a diverse background. DenseNet121 among the other networks achieved the highest validation accuracy of 87.86% and provided a good generalization and reliability in real-life applications. While MobileNetV2 performs well in training, it struggles with overfitting, and InceptionV3 maintains a good balance between precision and recall. Overall, these findings support and highlight the possibility for AI-facilitated ASL recognition to break communication barriers and help with assistive technology, providing inclusion all over the world with individuals reliant on sign language in communication.