Publication

Abstract : The social significance of recognizing sign language is in the provision of communication means to persons with hearing or speech disorders. In this work, a promising sign language recognition scheme based on the deep learning technique is developed. This system utilizes MobileNetV1 which is a low latency and resource-friendly convolutional neural network architecture modified by transfer learning. For training the model, a 21-class custom sign dataset was used, along with different transformations to the collected sign data. the MobileNetV1 based proposed model yields an accuracy of 97.14 percent on the test set to demonstrate the model capability. Additionally, weighted precision of 0.97, weighted recall of 0.97, and F1 score of 0.97 also supports the ability of the proposed model. Comparative analysis of MobileNetV1 in conjunction with MobileNetV2 and DenseNet establishes the former's effectiveness in high accuracy performance with moderate computational complexity, ideal for real-time usage. Training of the system was carried out by employing various methods including data augmentation and regularization, to develop good generalization and high robustness of the system. This paper also elaborates on new ideas of deep learning models in sign language recognition to real and demanding application into practices for accessibility. This work underlines that the lightweight architectures are crucial to overcome in the conditions of resource limitation and opens the way to the further evolution of the assistive technologies.