Comparative Analysis of Activation Functions in CNN for Digit Sign Language Classification

Sign language (SL) is a language that presented through hand signs and gestures as a form of communication for the deaf community. Due to the lack of SL understanding in society, there is still a communication gap between deaf individuals and society. Deep learning techniques, particularly of convolutional neural network (CNN) have shown a great contribution in SL classification. However, the selection of activation functions in the network is underexplored of its convergence and generalization impact to the CNN. Therefore, this study designs a custom 7-layered CNN to investigate the impact of activation functions on CNN performance for digit SL classification. A comparative analysis is conducted on different activation functions such as rectified linear unit (ReLU), hyperbolic tangent (Tanh) and Sigmoid. The aim of this study is to evaluate and select the suitable activation function for enhanced CNN-based SL classification system. The results of this study demonstrate that ReLU contributes the highest classification training accuracy of 83.11% and validation accuracy of 80.53%, 114.74 minutes training time and test error of 37.33% that highlights its effectiveness on convergence and feature extraction in CNN.