Abstract: The fashion industry plays an important role in the global economy, and its GDP scale shows a growing trend, currently accounting for approximately 2% of the global GDP. Fashion style classification can help consumers better understand and choose fashion products that suit their preferences, making it easier for them to find the styles and brands they are interested in. At the same time, fashion style classification also plays an important role in fashion research and design, providing inspiration and reference for designers. In addition, for fashion brand companies and retailers, the precise classification of fashion styles helps to better understand market demand and consumer preferences, thereby adjusting product categories and promotion strategies. Manual fashion style classification has the characteristics of subjectivity, diversity, variability and regionality, which can easily lead to errors in classification results. Therefore, it is important to improve the objectivity and accuracy of classification by using technical means, such as artificial intelligence. 
In response to the above problems, this paper studied a fashion style image classification method that combines transfer learning and ensemble learning. Firstly, based on the FashionStyle14 data set, duplicate or invalid images were filtered out to construct a fashion style image data set. Secondly, pre-trained models such as EfficientNet V2, RegNet Y 16GF and ViT Large 16 were used for fine-tuning training to generate new models to achieve fashion style image classification with a single deep learning model. Thirdly, the new model was tested and its classification performance was evaluated according to the evaluation indicators, and a good deep learning model was selected. Fourthly, an ensemble learning method based on voting, weighted integration and stacking was built to perform combined predictions on the above models, so as to improve the generalization ability and stability of the model. Finally, the ensemble learning method with the best performance was selected to classify fashion styles. By using the above method, not only advanced algorithms were applied in artificial intelligence to fashion style classification, but also the effectiveness of the method in classifying various fashion styles was verified, and some similar fashion styles were discovered, which provided data support for further refinement of fashion style classification. It is found that compared with the deep learning model based on a traditional convolutional neural network, the deep learning model based on the self-attention mechanism shows better recognition ability in fashion style image classification and recognition; compared with a single model, the commonly used integrated learning method can effectively improve the accuracy of fashion style image classification and recognition; the three styles of femininity, girly style and maiden style are similar, and it is easy to confuse between rock style and street style; cross-domain images are more likely to lead to recognition failure of fashion style images.
In future research work, similar styles can be analyzed more deeply based on methods such as fashion attributes to improve the performance and classification accuracy of the model. In addition, in-depth research on cross-domain fashion style image classification and recognition issues is needed to explore more effective solutions.

Key words: fashion style, transfer learning, ensemble learning, ViT model, image classification

摘要： 针对服装风格人工分类受主观性、地域等因素影响而造成的分类错误问题，研究了一种基于人工智能的服装风格图像分类方法。首先，在FashionStyle14数据集基础上筛除重复或无效图像，构建服装风格图像数据集；然后，采用迁移学习方法，对EfficientNet V2、RegNet Y 16GF和ViT Large 16等模型进行微调训练，生成新模型，实现基于单个深度学习的服装风格图像分类；最后，为进一步提高图像分类的准确性、可靠性和鲁棒性，分别采用基于投票、加权平均和堆叠的集成学习方法对上述单个模型进行组合预测。迁移学习实验结果表明：基于ViT Large 16的深度学习模型在测试集上表现最佳，平均准确率为77.024%；集成学习方法实验结果显示，基于投票的集成学习方法在相同测试集上平均准确率可达78.833%。该方法为解决服装风格分类问题提供了新的思路。

关键词: 服装风格, 迁移学习, 集成学习, ViT模型, 图像分类