EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC 4.0

Design and Implementation of a Cross-Cultural Pattern Classification System Based on ResNet18

Alina Andreeva, Zhenning Li, Wenjie Tang, Chaofan Yue, Xinlei Cao, Xiang Zhang, Xiangyu Li

Journal of Advances in Engineering and Technology · 13.05.2026 · 3 (2)

DOI: 10.62177/jaet.v3i2.1367 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Advances in Engineering and Technology, online erschienen am 13.05.2026. Verfasst von Alina Andreeva und 6 weiteren Autor:innen. Die Arbeit nennt 32 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The automatic recognition of traditional cultural patterns from China, Japan, and Korea poses a challenging cross-cultural image classification problem, compounded by high intra-class stylistic diversity, inter-cultural iconographic overlap, and the scarcity of annotated training data. We propose a lightweight transfer learning framework in which a ResNet18 backbone pre-trained on ImageNet is fully frozen and a compact three-class fully connected head comprising only 2,565 trainable parameters is appended and optimized for the target task, together with a stochastic data augmentation pipeline combing random resized cropping, horizontal flipping, rotation, and Colour jitter. A purposebuilt benchmark of 750 balanced images spanning Chinese, Japanese, and Korean traditional patterns is constructed and partitioned following a 70:15:15 ratio. Comprehensive evaluation against five competing architectures (VGG16, MobileNetV2, EfficientNet-B0, DenseNet121, and a from-scratch SimpleCNN baseline) under identical training conditions demonstrates that the proposed method achieves 87.61% test accuracy and an F1-score of 0.8746, outperforming SimpleCNN by 8.85 percentage points while requiring over four orders of magnitude fewer trainable parameters than VGG16. Controlled ablation experiments reveal that data augmentation contributes the largest single performance gain of 22.00 percentage points, followed by ImageNet pre-training at 11.61 percentage points. Per-class analysis shows that Korean patterns are most reliably discriminated (F1 = 0.900), Japanese patterns achieve the highest recall (0.949), and Chinese patterns pose the greatest challenge (recall = 0.765) owing to their broad stylistic range and historical iconographic diffusion into neighboring traditions. These findings establish that a frozen-backbone transfer learning strategy with targeted augmentation provides an efficient and practically deployable solution for cross-cultural heritage pattern recognition under realistic computational constraints.

Zitieren

APA

Andreeva, A., Li, Z., Tang, W., Yue, C., Cao, X., Zhang, X., & Li, X. (2026). Design and Implementation of a Cross-Cultural Pattern Classification System Based on ResNet18. Journal of Advances in Engineering and Technology, 3(2). https://doi.org/10.62177/jaet.v3i2.1367

Vancouver

Andreeva A, Li Z, Tang W, Yue C, Cao X, Zhang X, et al. Design and Implementation of a Cross-Cultural Pattern Classification System Based on ResNet18. Journal of Advances in Engineering and Technology. 2026;3(2). doi:10.62177/jaet.v3i2.1367

BibTeX
@article{Andreeva2026,
  author = {Andreeva, Alina and Li, Zhenning and Tang, Wenjie and Yue, Chaofan and Cao, Xinlei and Zhang, Xiang and Li, Xiangyu},
  title = {Design and Implementation of a Cross-Cultural Pattern Classification System Based on ResNet18},
  journal = {Journal of Advances in Engineering and Technology},
  year = {2026},
  volume = {3},
  number = {2},
  doi = {10.62177/jaet.v3i2.1367}
}
Kostenlos starten →