Using an enhanced U-Net architecture to segment intricate motifs from heritage-based wood carvings

Wood carving is an important form of intangible cultural heritage with rich artistic and historical value. Accurate motif segmentation is essential for digital analysis and preservation, but this task remains difficult because wood-carving images often contain dense textures, fine details, and complex backgrounds. To address this problem, this study proposes an Enhanced U-Net (EU-Net) for wood-carving motif segmentation. The model combines residual connections with a pixel-wise adaptive loss weighting strategy to improve feature extraction and sensitivity to difficult pixels while keeping computational cost relatively low. A dedicated dataset of digitized wood carvings was constructed in collaboration with a nationally recognized Intangible Cultural Heritage inheritor. Comparative experiments show that EU-Net achieves better results than U-Net, Attention U-Net, and Dense U-Net. Per-class IoU, Dice, and 5-fold cross-validation further support its effectiveness and stability. The study also examines the potential of this model for future use in a digital wood-carving collection platform.