Analysis and research on traditional Zhuang costume patterns based on an improved MobileNetV3

The traditional Zhuang costumes are characterized by their intricate design and cultural symbolism, but are becoming more and more endangered with urbanization and cultural homogenisation. This paper suggests a lightweight model that integrates MobileNetV3-small, the YOLO framework, and SimAM attention mechanism to better extract features to support their recognition and digital preservation. The model applies depthwise separable convolutions and multi-scale attention to have a better look at complex textile textures. When trained on a custom dataset of ~2,000 images, it is 8% 10% to the base model in terms of mAP@0.5, and ~9 GFLOPs and 3,478,258 parameters cheaper. This technique is very precise and has great efficiency and can be applied on a resource limited devices, it is also useful in maintaining old ethnic trends.