Research on automatic image classification and feature extraction of ethnic minority costumes in northern Guangdong based on deep learning

With the rapid development of deep learning technology, image classification and feature extraction have been widely used in the study of cultural heritage and ethnic costumes. The costumes of ethnic minorities in northern Guangdong have rich cultural connotations and unique design styles, but due to the low efficiency and strong subjectivity of traditional classification methods, they have not been effectively automated. Therefore, how to achieve efficient and accurate clothing classification has become a key issue that needs to be solved urgently.This study proposes an automatic classification and feature extraction method based on deep learning to improve the image recognition accuracy of ethnic costumes in northern Guangdong. We selected two mainstream deep learning architectures, ResNet50 and Vision Transformer (ViT), and compared their performance. Through transfer learning technology, we trained these two models on a small sample dataset and optimized them using the cross entropy loss function and Adam optimizer. Experimental results show that ViT outperforms ResNet50 (accuracy 92%, F1 score 0.91) in both classification accuracy (94%) and F1 score (0.93). In addition, the Grad-CAM technology was used to visualize the feature extraction results of the model, revealing the model's focus on key areas in clothing images, such as patterns, colors, and shapes.This study not only demonstrates the potential of deep learning technology in the classification of ethnic clothing images, but also proposes an effective multi-model comparison method, which promotes the development of cultural heritage digital processing technology. Through this method, more accurate and efficient technical support can be provided for the automated research and protection of ethnic minority clothing in northern Guangdong.