Comparison of Kernel Convolutional Neural Network in Lampung Script Word Recognition and Transliteration

Abstract

The study aims to create a system that can recognize and transliterate Lampung script image data and compare the Convolutional Neural Network (CNN) kernel to the Lampung script word recognition and transliteration system. The Lampung script recognition and transliteration system with the CNN learning model is implemented using the python 3.9.4 64 bit programming language, with a stride of 1 for convolution and 2 for pooling, the kernel size variations used are 2x2, 3x3 and 5x5 which are applied crosswise for feature extraction of the convolution and pooling processes. The 3x3 convolution kernel type and 3x3 pooling kernel showed the best performance in transliterating and recognizing Lampung script words with a test accuracy of 0.9 and a small test result data inequality, which is 2/10 or 0.2. The 3x3 Kernel Size shows ideal conditions for use, especially when the image features used have very few differences in features.