Comparing Anime Character Classification with MobileNetV2, ResNet50V2, and Xception

Abstract

This study compares three pretrained CNN models including MobileNetV2, ResNet50V2, and Xception for anime character classification using transfer learning. We apply these models to anime images, which are challenging due to their high visual variation and stylized appearance. We also include dropout regularization to improve model generalization and reduce overfitting. We propose a simple evaluation framework using validation, test, and external internet datasets. The results show that all models perform well on anime character classification. ResNet50V2 with dropout achieves the best and most stable performance, reaching 96.36% validation accuracy, 96.00% test accuracy, and 96.53% on the internet dataset. Dropout improves performance across all models, especially for Xception, which is more prone to overfitting. MobileNetV2 delivers slightly lower accuracy but offers much higher efficiency, making it suitable for lightweight applications. We explore the main sources of classification errors and find that most mistakes occur in visually similar characters with high stylistic variation. We conclude that ResNet50V2 with dropout is the most reliable model for this task, while MobileNetV2 is the best option for efficient deployment. Future work can improve performance further using better augmentation strategies and ensemble learning.