Cultural heritage image classification using transfer learning for feature extraction: a comparison

Abstract

Image classification in the domain of cultural heritage becomes extremely important with the development of digitization practices. This study aims to analyze how classification performance on the small dataset representing cultural heritage changes depending on the feature extraction method. The dataset comprised of 150 images belonging to three classes: (i) archaeological sites, (ii) frescoes, and (iii) monasteries. Five transfer learning architectures were used to extract the features from images, while classification was per-formed using four traditional machine learning algorithms, mainly Random forest, Naïve Bayes, Decision tree, and Multilayer perceptron classifier. The results suggest that Random forest and Multilayer perceptron are the most suitable algorithms for classification of cultural heritage images, especially when used in combination with the DenseNet121 pre-trained architecture. Naïve Bayes also performed well, with a maximum accuracy of 100% obtained when features are extracted using EfficientNetB0. However, the Decision tree algorithm reached only moderate performance.