Improving Ancient Language Classification with Deep Learning and Traditional Models for Class Imbalance

Abstract

The paper suggests a hybrid framework of the ancient language image classification, which combines deep transfer learning with ensemble machine learning to overcome the issue of imbalance in the classes in low-resource settings. They use a pre-trained Deep Visual Feature Extractor (DVFE) that is built on the ResNet-50 architecture to produce the high-level visual representations of the Ancient Language Images (ALI) dataset that consists of eight historical scripts. In order to reduce moderate levels of imbalance in the deep feature space, imbalance-aware resampling algorithms, such as SMOTE and ADASYN are only used on the training set. They are then used to extract features which are then further classified with the help of Random Forest and LightGBM models with hyperparameters being optimized via cross-validation. On a held-out test set, experimental results show strong and balanced performance with an overall accuracy of 0.89, macro-precision of 0.88, macro-recall of 0.87, macro-F1 of 0.88 and per-class values of AUC of 0.88 to 0.97. The results show that the ability to correct the imbalance on the feature level allows to increase the fairness between the majority and minority language classes and retain good discriminative power. This comparison also supports the fact that the proposed DVFE-ensemble system yields better performance as compared to the traditional methods used to deal with imbalance on the raw images level. On the whole, the research project demonstrates the usefulness of the deep feature extraction and imbalance-conscious ensemble learning as a tool of ensuring a high level of reliability in the ancient languages classification, which will further the development of the digital heritage preservation systems in the conditions of limited amounts of information.