News Summary

• Southwest traditional crafts face recognition challenges due to intricate textures, variable scales, and a deficiency in specialized data.
• This study proposes the Scale-Wise Convolutional Refinement Attention Fusion Network (SW‑CraftNet), which integrates multi-scale features and dual attention mechanisms.
• Experimental results indicate that the model achieves 96.83% accuracy on the Chinese Intangible Cultural Heritage Image Dataset.
• This performance outperforms mainstream architectures such as ResNet50, showing competitive performance in complex categories like Miao embroidery and lacquerware.