Research on a prompt learning method for intangible cultural heritage art image classification

To address the issues of prolonged processing time, low efficiency, and high data complexity in the classification of Chinese intangible cultural heritage (ICH) artworks, this paper proposes a context-based text prompt tuning strategy based on a pre-trained vision-language model. This approach introduces trainable context optimization soft prompts, enabling the model to quickly adapt to downstream classification tasks under limited sample conditions, thereby effectively reducing training time and improving convergence speed. Specifically, the proposed method integrates text features generated by the soft prompts with the original features of the pre-trained vision-language model through an attention mechanism, and optimizes the embedded representations via a contrastive loss function. This mechanism significantly reduces the embedding discrepancy between the two types of features, preventing the model from overfitting to visible base categories and enhancing its generalization ability to unseen classes. Moreover, the retention of original features helps mitigate catastrophic forgetting during training, ensuring high classification accuracy even under few-shot conditions. Experimental results demonstrate that the proposed method improves classification accuracy by 1.79%, enhances generalization by 10.4%, and maintains low computational cost.