Image preprocessing analysis in handwritten Javanese character recognition

Abstract

The handwriting produced by each person is unique, so each person has a different stroke, even though they write the same letter. Handwritten Javanese is an exciting topic to study, in addition to scientific purposes and preserving Indonesian culture. The Javanese character image dataset is aksara Jawa: aksara Jawa custom dataset from the Kaggle database consists of 2,154 train data and 480 evaluation data. This research proposed to analyze the impact of some preprocessing methods in recognizing handwritten Javanese characters. The preprocessing methods are dilation, skeletonization, and noise reduction. The first process is segmentation for region of interest (ROI) extraction, then various preprocessing is used, and finally, the recognition step neural network (NN) to measure the effectiveness of the preprocessing method. The experiment shows that all preprocessing methods (dilation, skeletonization, and noise reduction) give excellent results, especially on the black background color, reaching 98% accuracy. Other experimental findings show that in any preprocessing combination, the black background accuracy is better than the white one.