Research on Similarity Matching of Dazu Rock Carvings in China Based on a Trimodal Fusion of pHash, SIFT, and MSRoVGG

Song Dynasty grottoes, represented by the Dazu Rock Carvings, exhibit pronounced artistic standardization and stylistic convergence. However, conventional similarity assessments of cultural relics remain heavily reliant on qualitative expert experience, which often results in inconsistent evaluations and lacks the quantitative rigor necessary for refined digital conservation and restoration. To address this, we propose a quantitative evaluation model based on trimodal feature fusion. The framework first employs the Perceptual Hash (pHash) algorithm for global structural screening of orthophotos, followed by the integration of the SIFT operator to extract high-dimensional texture features and perform local keypoint matching. Finally, an improved MSRoVGG deep learning architecture is incorporated to achieve deep feature verification at the semantic level. Using the Seven Buddhas of the Present Kalpa in the Dazu Rock Carvings as a case study, the results demonstrate a 100% recognition rate for identical statue types within the same niche. Furthermore, the model successfully discriminates between different iconographic types, effectively capturing semantic variations in artistic styles and vestimentary features between Buddhist and secular figures. These findings confirm the model's robust capability in typological classification and stylistic differentiation.