Deep Learning-Based Recognition of Miao Ethnic Costumes via YOLOv5s: A Step Toward Digital Cultural Preservation

Abstract Miao ethnic costumes, celebrated for rich diversity, intricate craftsmanship, and distinctive patterns, represent an important aspect of China's cultural heritage and the broader realm of intangible cultural heritage. In response to the growing need for digital preservation, this study proposes a deep learning-based approach to effectively recognize and document Miao costumes. While traditional object detection methods face challenges such as high computational costs and limited analytical capacity, the YOLOv5s framework offers automatic feature extraction and improved scalability. However, its standard form struggles to adequately focus on critical visual features, reducing recognition performance accuracy. To overcome this, we introduce the YOLOv5s-SED model, which incorporates a Squeeze-and-Excitation (SE) attention mechanism and deformable convolution (DCNv2) into YOLOv5s to enhance feature representation and improve the detection of fine details. A dedicated dataset of 4,468 annotated images was compiled, and the model was refined through hyperparameter tuning and comparative experiments. The results demonstrate notable performance gains, with precision increasing from 97.6–98.1%, recall from 99.3–99.8%, and mean Average Precision (mAP) from 70.7–71.5%. These outcomes highlight the model's strong generalization ability in complex environments and its potential to support the digital preservation and promotion of Miao ethnic costumes.