Comparison of Multi-Label Learning Approaches for Pie Chart Image Classification Using Deep Learning

Birant D., Kösemen C.

Journal of Artificial Intelligence and Data Science, cilt.1, sa.2, ss.116-124, 2021 (Hakemli Dergi)

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 1 Sayı: 2
  • Basım Tarihi: 2021
  • Dergi Adı: Journal of Artificial Intelligence and Data Science
  • Derginin Tarandığı İndeksler: Other Indexes
  • Sayfa Sayıları: ss.116-124
  • Dokuz Eylül Üniversitesi Adresli: Evet

Özet

A pie chart is a powerful and circular information graphic used to display numerical proportions to the whole. However, the properties of pie charts cannot be directly noticed by machines since they are usually in an image format. To make a pie chart classifiable by machines, this paper proposes a novel solution using deep learning methods. This study is original in that it automatically and jointly classifies charts in terms of two respects: shape (pie or donut) and dimension (2D or 3D). This is the first study that compares two multi-label learning approaches to classify pie charts: binary-class-based convolutional neural networks (BCNN) and multi-class-based convolutional neural networks (MCNN). The experimental results showed that the BCNN model achieved 86% accuracy, while the MCNN model reached 85% accuracy on real-world pie chart data.