Authenticity identification of Qi Baishi's shrimp painting with dynamic token enhanced visual transformer

Abstract

Automatic recognition of Chinese ink paintings’ authenticity is still a challenging task due to the high similarity between genuine and fake paintings, and the sparse discriminative information in Chinese ink paintings. To handle this challenging task, we propose the Dynamic Token Enhancement Transformer (DETE) to improve the model’s ability to identify the authenticity of Qi Baishi’s shrimp paintings. The proposed DETE method consists of two key components: dynamic patch creation (DPC) strategy and dynamic token enhancement (DTE) module. The DPC strategy creates patches with different sizes according to their contributions, forcing the network to focus on the important regions instead of meaningless ones. The DTE module gradually enhances the association between the class token and most impact tokens to improves the performance eventually. We collected a dataset of authenticity identification of Qi Baishi’s shrimp paintings and validated our method on this dataset. The results showed that our method outperformed the state-of-the-art methods. In addition, we further validated our method on two public available painting classification datasets WikiArt and ArtDL.