Javanese Script Classification with YOLO26 and Convolutional Filter Visualization

Abstract

Javanese script is a cultural heritage whose use is narrowing, while automatic recognition is held back by visual similarity between characters. This study evaluates the YOLO26 architecture on Javanese script classification, records the change in convolutional kernel weights across the eleven architecture blocks, and tracks prediction probabilities on visually similar characters. Secondary data of 2,659 Nglegena script images from 20 classes came from Kaggle. Each image passed through Otsu's binarization, diacritic-safe bounding, resizing to 224x224 pixels, and letterboxing; the training partition was then expanded through augmentation. The YOLO26 nano variant was trained in PyTorch with the AdamW optimiser and stopped at the tenth epoch through early stopping. Kernel weights were read from every convolutional module each epoch. Testing on 25 images produced 96% accuracy with a macro average F1-score of 0.9733, and 18 of the 20 classes were recognised without error. Blocks 1 to 3 recorded edges, corners and curves at larger weight values, blocks 4 to 9 produced denser representations with weights closer to zero, block 10 concentrated on class-separating areas, and block 11 converted them into class scores. By the tenth epoch most classes stayed above 99%, while 'ha' reached 85%.