Classifying Intangible Cultural Heritage Images in the Mekong Delta

Our study aims to classify images of Intangible Cultural Heritage (ICH) in the Mekong Delta, Vietnam, a region well-known for its rich cultural diversity. To achieve this purpose, we have built a comprehensive dataset consisting of images from 17 distinct ICH categories, including various forms of traditional practices, performances, and cultural expressions. We manually annotated images of this image dataset. Initially, we fine-tuned recent pre-trained network models, including VGG16, DenseNet, and Vision Transformer (ViT), for classifying our dataset. After that, we trained Logistic Regression (LR) models, called fusing models, which fuse not only various visual features extracted from deep networks but also the output of deep networks to improve the classification accuracy. Futhermore, we develop a local training algorithm to perform the classification task from visual features extracted by ViT from images. Our comparative study of the classification performance on the 17-category ICH image dataset shown that our fusing models improve the classification correctness compared to any single fine-tuned deep learning model. The first fusing model, which applies LR using the visual features extracted from VGG16, DenseNet, and ViT, achieved an accuracy of 66.76\%. The second fusing model, which applies LR directly on the outputs of VGG16, DenseNet, and ViT, yielded an accuracy of 66.49\%. Specifically, our local training algorithm using only ViT features achieved a best overall classification correctness of 68.22\%. These results illustrate the effectiveness of our fusion approach in leveraging the strengths of multiple models and local training algorithm to improve overall classification correctness.