GCP-Mamba: fusion of geometric and compositional priors in state space models for Thangka image captioning

Thangka, a distinct cultural heritage of Tibet, is characterized by rigorous symmetric layouts, concentrated core semantics, and a unique "centric-radiant" composition. However, existing Mamba-based image captioning models lack domain-specific prior knowledge, making it difficult to generate precise and culturally compliant descriptions. To address this, we propose GCP-Mamba (Geometric and Compositional Priors Mamba), a state space model that integrates geometric and compositional priors. Specifically, we design a Symmetry-Aware Fusion (SAF) module to explicitly model geometric symmetry, thereby calibrating spatial orientation logic and enhancing feature robustness. Furthermore, we introduce a Compositional Priors State Space Model (CP-SSM), which concurrently disentangles the fine-grained entity associations within the Principal Deity region and the macro-level "inside-out" layout. Experimental results on the D-Thangka dataset demonstrate that GCP-Mamba achieves a CIDEr score of 895.7 and a BLEU-4 of 94.4, outperforming baseline models by 34.18% and 3.20%, respectively. These results validate the effectiveness of our proposed modules in capturing the complex structures and microscopic details of Thangka imagery, providing a novel technical path for the intelligent protection of digital cultural heritage.