Analisis Pengaruh Peningkatan Kualitas Citra Berbasis Histogram terhadap Kinerja ResNet-18 pada Klasifikasi Karakter Katakana

Abstract

Handwritten Katakana character classification remains a challenging task due to variations in individual writing styles and suboptimal image quality, such as low contrast and unclear character details. These conditions can affect the model's ability to extract discriminative features and accurately recognize characters. This study aims to analyze the impact of histogram-based image enhancement techniques, namely Histogram Equalization (HE), Adaptive Histogram Equalization (AHE), and Contrast Limited Adaptive Histogram Equalization (CLAHE), on the performance of the ResNet-18 model for Katakana character classification. The ETL5 dataset, consisting of 9,200 handwritten character images across 46 character classes, was used in this study. The dataset was divided into training and testing sets with an 80:20 ratio. To increase data diversity and reduce the risk of overfitting, data augmentation using rotation and translation was applied to the training set. The classification process employed a transfer learning-based ResNet-18 model initialized with ImageNet pretrained weights, where only the fully connected layer was fine-tuned using the Adam optimizer for 50 epochs. The experimental results show that CLAHE achieved the best performance with an accuracy of 97.01%, while HE and AHE obtained accuracies of 83.15% and 76.79%, respectively. These findings indicate that CLAHE is more effective than HE and AHE in improving the classification performance of handwritten Katakana characters using ResNet-18, resulting in more accurate predictions. This study contributes by providing a comparative analysis of the effects of three histogram-based image enhancement methods on the performance of ResNet-18 for handwritten Katakana character classification.