Sign language provides communication bridge for spoken language and visual gesture. However, Automatic Recognition of Sign Language is a complicated system due to the diversity of various sign languages and difference in hand movement, body gesture, and facial expression. Most often convolutional neural network models are employed to identify sign language and activation function is a crucial part of the CNN model’s hierarchical structure because of its non-linear characteristics. Seven frequent non-linearity functions—LeakyReLu, PReLu, ReLu6, ReLu, Swish, HardSwish, SELU, and Mish—have been examined and assessed for recognition of sign language using Tensorflow library. This study compares the result of convolutional neural network (CNN) against various non-linearity functions. Findings of the experiments indicate that CNNs based on the Mish activation function with an accuracy of 98.63 perform better than CNNs using the other activation functions.