Utilizing Convolutional Neural Network for Learning Web-Based Braille Letter Classification System

Abstract

This paper aims to facilitate prospective teachers and people who want to learn braille letters. The system designed is a website that will classify braille letters using the convolutional neural network (CNN) method with the activation functions used, namely ReLU and Softmax. In this research, the input is an image of braille letters with grayscale elements. The output of the data is a regular alphabet letter. Most of this research data consists of training and testing data, which is 2,722 pieces. The accuracy results obtained in the data training process using Max Pooling and epoch 30 for data is 92.15%, epoch 50 is 94.58%, and for training data with epoch 100 is 96.64%. The test results using the system produce an accuracy value of all braille letter image data of 92.30%. Furthermore, for better system development, it is recommended to use hyperparameter tuning to minimize classification uncertainty in braille letter images.