AI framework classifies avian bones with 86% accuracy
A proof-of-concept multimodal framework integrating convolutional neural network-based image analysis with osteometric measurements achieves 86% accuracy in classifying bird bones. Using over 10,000 images from multiple museum and research collections, the system performs skeletal element identification and family-level taxonomic classification. Images are automatically segmented via a two-stage pipeline combining BiRefNet and SAM2. Visual features from a pre-trained EfficientNet_V2_S backbone are fused with standardized morphometric data through a feature-level architecture. The research, published on arXiv, demonstrates AI's potential in zooarchaeology, a field where its application has been limited.
Key facts
- Multimodal framework combines CNN image analysis with osteometric measurements
- Dataset includes more than 10,000 images from multiple museum and research collections
- Two classification tasks: skeletal element identification and family-level taxonomic classification
- Image segmentation uses two-stage pipeline with BiRefNet and SAM2
- Visual features extracted with pre-trained EfficientNet_V2_S backbone
- Feature-level multimodal architecture fuses visual and morphometric data
- Model achieved 86% accuracy on test set for bone-type classification
- Published on arXiv as a proof-of-concept for AI in zooarchaeology
Entities
Institutions
- arXiv