Non-destructive wood species identification in cultural heritage artefacts: Multi-angle CT imaging combined with AI-enhanced recognition

Non-destructive wood species identification is a critical requirement in cultural heritage conservation, yet traditional thin-section anatomical methods demand destructive sampling unacceptable for irreplaceable artefacts. This study presents a practical, open-source framework combining multi-angle X-ray micro-computed tomography (Micro-CT) with an enhanced YOLOv8 deep learning model for direct adoption by museum and conservation practitioners. CT images acquired from three orthogonal sections (transverse, radial, tangential) capture complementary anatomical features—vessel distribution, growth ring patterns, and density variations—enabling species discrimination in traditional furniture, carvings, and musical instruments without physical contact. A Multi-Scale Dynamic Weighting (MSDW) mechanism—the core architectural innovation of this work, inspired by dynamic convolution attention—adaptively extracts both large-scale structural patterns and fine-grained anatomical detail from CT images. Physics-informed feature analysis identifies mean density and vessel distribution as the primary discriminants for species classification, followed by growth ring contrast and anisotropy indices. The optimised V8NGB model achieves 97.0% mAP50 across 17 wood species central to Chinese cultural heritage—a 9.4% gain over baseline YOLOv8—at only 6.5 ms inference time, meeting the demands of real-time conservation workflows. All trained models and source code are openly available at https://github.com/liwei0214/wood-CT-heritage .