A stitch recognition based on ResNet and an attention mechanism for Cantonese Embroidery

Stitches are the defining feature of Cantonese Embroidery, an intangible cultural heritage of China. Their diversity and high flexibility pose significant challenges for automatic recognition. A new framework for the recognition of Cantonese Embroidery stitches in real-world embroidery scenarios is proposed. First, the stitch regions are segmented and extracted from embroidery images using SAM2, and a dataset is then established. Afterward, the Squeeze-and-Excitation module is refined and incorporated into a CNN structure, leading to an enhanced Squeeze-and-Excitation Residual Neural Network for Cantonese Embroidery (SE-ResNet-CE) stitch recognition model. Through this design, joint attention to global and local stitch features is achieved. The experimental results reveal that the proposed model achieves a classification accuracy of 98.81% and a Macro-F1 score of 98.3% on the dataset. This method effectively extracts fine-grained features of stitches, providing technical support for the digitization of embroidery.