Deep CNN for Static Indian Sign Language Digits Recognition

Sign language recognition (SLR) is a significant solution for the hearing and speech disabled to connect with the people. However, SLR system faces complexities such as low accuracy, overfitting, hand occlusions, and high interclass similarities. In this paper, a deep learning-based Convolution Neural Network model is proposed for Sign language recognition to address the issues. Our model uses Indian Sign Language dataset which comprises 10 class with a total of 2072 static digit gestures ranging between 0 to 9. Each class has 207 images. The proposed model generated desired outcome and the results are evaluated with varied optimizers such as Adam, RMS Prop, Stochastic gradient descent (SGD) optimizers. CNN model with SGD achieved training and validation accuracy of 99.72% and 98.97% respectively. The training and validation loss were comparatively minimum for our model. Further, the performance evaluation of the proposed model was analyzed based on precision, recall, F-score value. Our method shows its effectiveness over other machine learning models with a recognition rate of 99%.