Analisis Komparatif Arsitektur CNN dan ViT dengan Variasi Augmentasi untuk Klasifikasi Karakter Mandarin

Abstract

Handwritten Chinese Character Recognition (HCCR) is a fundamental task in computer vision that poses significant challenges due to the large number of character classes and high variability of writing styles. This study aims to perform a systematic comparative evaluation between Convolutional Neural Network (ResNet-50) and Vision Transformer (ViT-B/16) architectures in handling these challenges under limited data scenarios. The research utilizes a specific subset of the CASIA-HWDB1.1 dataset, consisting of 10 fine-grained character classes selected for their visual similarity. To ensure robust evaluation a 5-Fold Cross-Validation method was employed across five experimental scenarios (Baseline, Geometric, Elastic Transform, Random Erasing, and CutMix). Experimental results demonstrate that ResNet-50 consistently outperformed ViT-B/16 in terms of accuracy and stability. The highest performance was achieved by ResNet-50 using Elastic Transform augmentation with a test accuracy of 90.52%, whereas the best performance for ViT-B/16 was achieved using Random Erasing at 88.69%. The study also reveals that ViT exhibits higher performance variability compared to CNN. These findings conclude that for HCCR tasks with limited data, CNN possess a superior inductive bias for capturing local stroke features while data augmentation must be tailored specifically to the architecture type to maximize performance.