Extraction of Xiuyan embroidery patterns based on vision transformer

Aiming at the pain points of low efficiency in manual extraction of Xiuyan Embroidery patterns (4-6 hours per piece) and insufficient accuracy of traditional models caused by texture noise (Dice coefficient ≤ 0.7), this study takes MobileViTLite, a lightweight branch of Vision Transformer, as the foundation. The model is optimized with multi-scale attention and noise suppression masks, and a standardized dataset is constructed using over 1,200 embroidery samples to focus on the full-process development of digital pattern extraction. Experimental results show that the optimized model achieves a Dice coefficient of 0.892 and an IoU of 0.867, which are increased by 20.1% and 19.9% respectively compared with the traditional U-Net. The inference speed on the Jetson Nano edge device is ≤ 42 ms per image, and the number of parameters is controlled within 5.85 million. This research provides a feasible engineering technical path for the digital protection and inheritance of intangible cultural heritage (ICH) embroidery, and plays a positive supporting role in promoting technological innovation and practical application in the field of traditional craft digitization.