Improved Javanese script recognition using custom model of convolution neural network

Abstrak

Handwriting recognition in Javanese script is not widely developed with deep learning (DL). Previous DL and machine learning (ML) research is generally limited to basic characters (Carakan) only. This study proposes a deep learning model using a custom-built convolutional neural network to improve recognition accuracy performance and reduce computational costs. The main features of handwritten objects are textures, edges, lines, and shapes, so convolution layers are not designed in large numbers. This research maximizes optimization of other layers such as pooling, activation function, fully connected layer, optimizer, and parameter settings such as dropout and learning rate. There are eleven main layers used in the proposed custom convolutional neural network (CNN) model, namely four convolution layers+activation function, four pooling layers, two fully connected layers, and a softmax classifier. Based on the test results on the Javanese script handwritten image dataset with 120 classes consisting of 20 basic character classes and 100 compound character classes, the resulting accuracy is 97.29%.