Research on the Classification Method of Ancient Paintings Based on Improved Swin Transformer

Quick and accurate identification of the author of ancient paintings holds significant importance for experts in the field of art authentication. Currently, the identification of ancient paintings relies on experts' knowledge of the relevant domain and their understanding of the background of the artworks, leading to issues such as strong human reliance and high difficulty in identification. In this context, we propose a method for ancient painting artist identification and classification based on an improved Swin Transformer model. Firstly, a dataset comprising 427 paintings from artists such as Qichang Dong, Kaizhi Gu, Zan Ni, Bohu Tang, Wei Wang, Daozi Wu, and Mengfu Zhao is utilized. Through image augmentation operations like rotation and mirroring, the dataset is expanded to 4060 paintings for this research. Secondly, five deep learning image classification models, namely VGG, ShuffleNet, AlexNet, MobileNet, and Swin Transformer, are selected for the study. Through comparative analysis, the Swin Transformer is identified as the optimal model. Finally, a CB module is incorporated into the Swin Transformer, and Poly-1 Loss is employed for ancient painting classification. Experimental results demonstrate improvements over VGG, ShuffleNet, AlexNet, and MobileNet models, the accuracy rates improved by 7.354, 20.663, 8.340, and 14.131 percentage points, respectively. When compared to the original Swin Transformer model, the accuracy rates improved of 3.606 percentage points.