EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-SA 4.0

Chart Classification of DocFigure Dataset using VGG-16 and EfficientNet-B3.

Reta L. Alsaoor, Hasan S. M. Al-Khaffaf

Zanin Journal of Science and Engineering · 2025

DOI: 10.64362/zjse.145 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Zanin Journal of Science and Engineering, 2025. Verfasst von Reta L. Alsaoor und 1 weiteren Autor:innen. Die Arbeit nennt 36 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-SA 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

This study investigates the automatic classification of chart images. We test two convolutional neural networks, VGG-16 and EfficientNet-B3, on the 28-class DocFigure dataset of approximately 33k images using two different methods: (i) original, unaugmented data and (ii) a class-balanced, augmented setting. Cross-entropy loss with Adam/AdamW, discriminative learning rates with scheduling, mixed precision, and best-checkpoint selection are used for training. Our contribution is a training strategy that starts from pre-trained CNNs (VGG-16 / EfficientNet-B3), fine-tunes of carefully selected layers, schedules learning rates, and applies weight decay and Automatic Mixed Precision (AMP), yielding higher chart type accuracy. VGG-16 achieves an overall classification accuracy of 82.65%, while EfficientNet-B3 attains an average classification accuracy of [A: 95.19%, P: 95.63%, R: 95.19%, F1: 95.36%] on the unaugmented split with σ ≈ 0.002 at the maximum of 5 train-test runs. In addition, VGG-16 accuracy improves to 89.20%, and EfficientNet-B3’s average accuracy after 5 train-test runs is [A: 94.95%, P: 95.06%, R: 94.94%, F1: 94.91%] with σ ≈ 0.001 on the augmented dataset. Compared to the original DocFigure baseline study (~88%), these results show that stronger pre-trained backbones and modern fine-tuning strategies together yield significant improvements without the need for custom architectures. Across both the original and augmented splits, EfficientNet-B3 remains the top performer, while augmentation substantially improves VGG-16’s classification accuracy.

Zitieren

APA

Alsaoor, R. L., & Al-Khaffaf, H. S. M. (2025). Chart Classification of DocFigure Dataset using VGG-16 and EfficientNet-B3. Zanin Journal of Science and Engineering. https://doi.org/10.64362/zjse.145

Vancouver

Alsaoor RL, Al-Khaffaf HSM. Chart Classification of DocFigure Dataset using VGG-16 and EfficientNet-B3. Zanin Journal of Science and Engineering. 2025. doi:10.64362/zjse.145

BibTeX
@article{Alsaoor2025,
  author = {Alsaoor, Reta L. and Al-Khaffaf, Hasan S. M.},
  title = {Chart Classification of DocFigure Dataset using VGG-16 and EfficientNet-B3},
  journal = {Zanin Journal of Science and Engineering},
  year = {2025},
  doi = {10.64362/zjse.145}
}
Kostenlos starten →