Imbalanced Classification of Cultural Heritage Images Using Deep Neural Networks
Juthatip Wisanmongkol, Taweesak Sanpechuda, Sambat Lim, La-or Kovavisaruch · 2024
We investigate the use of deep learning in cultural heritage image classification when the number of samples is not uniformly distributed among the classes. Two intuitive methods, data resampling and transfer learning, are considered to address the impact of the imbalanced dataset. Results show that classification performance can be improved by using the two methods. Compared with the baseline convolutional neural network, the overall F1 scores computed over all classes show slight improvements of 2.67% and 4.56% for the data oversampling and transfer learning methods. However, when the per-class F1 scores are considered, the upgrades can be as high as 64.77% and 67.35% for the data oversampling and transfer learning methods, respectively.