Leveraging Model Soups to Classify Intangible Cultural Heritage Images from the Mekong Delta

Researchers developed a new AI framework combining CoAtNet architecture with model soups technique to classify Intangible Cultural Heritage images from the Mekong Delta. The approach achieved 72.36% accuracy on the ICH-17 dataset, outperforming traditional models like ResNet-50 and ViT by reducing variance and improving generalization in low-resource settings.

• →Model soups technique averages checkpoints from training to reduce variance without increasing inference costs.
• →CoAtNet architecture captures both local and global patterns through fusion of convolution and self-attention mechanisms.
• →The framework achieved state-of-the-art 72.36% top-1 accuracy on ICH-17 dataset with 7,406 images across 17 classes.
• →Model soups selects geometrically diverse checkpoints, providing better generalization than traditional ensemble methods.
• →The approach addresses key challenges in cultural heritage classification including limited data and high visual similarity between classes.