Abstract

Gesture is one type of communication by forming an object such as letters or numbers on the hand to convey a message or information, one of which is the number gesture on the hand which has many types with a different patterns for each movement of the numbers formed. One solution that can be done to perform numeric gesture recognition on a computer is to use the Convolutional Neural Network (CNN) method with AlexNet and LeNet architectures. This study uses a numeric gesture image dataset that was previously carried out in the pre-processing stage consisting of threshold and resize. The research was conducted using 2 pooling layers, namely Average Pooling and Max Pooling and then using an optimizer, namely SGD, RMSprop, and Adam. Based on the test results obtained in this study, the use of the AlexNet architecture with Average Pooling and the RMSprop optimizer resulted in an overall accuracy and f1-score of 99.45% and the use of the LeNet architecture with Average Pooling and the RMSprop optimizer resulted in an overall accuracy and f1-score of 99.49%. Overall the use of Average Pooling with the RMSprop optimizer gets the best level of accuracy compared to other tests.