FCSRS: research on Thangka image captioning model combining feature compression supplement and semantic re-learning

Abstract

To address the challenges associated with the traditional manual interpretation of Thangka, which demands extensive prior knowledge and is time-consuming, this paper proposes a deep learning-based automatic captioning model for Thangka images, aiming to achieve precise visual feature extraction and deep semantic modeling. To overcome the limitations of existing methods in insufficient detail feature extraction and difficult semantic understanding of Thangka, a Feature Compression Supplement and Re-learning Semantic Thangka image captioning model (FCSRS) is proposed. Experimental results demonstrate that on our self-constructed Thangka dataset, this model achieves improvements of 17.1%, 20.7%, 17.3%, and 215.8% in BLEU-1, BLEU-4, METEOR, and CIDEr metrics, respectively, compared to the baseline model, significantly enhancing caption accuracy and semantic completeness.