Robust vision language semantic alignment for fine grained cultural heritage retrieval

With the growing demand for the digital preservation of cultural heritage, achieving precise alignment between image and textual information in complex environments has become a critical challenge in the fields of computer vision and artificial intelligence. Existing methods often encounter difficulties when processing cultural heritage data characterized by substantial noise and strong domain specificity, leading to insufficient cross-modal semantic alignment and limited robustness. This paper proposes Cross-Domain Vision-Language Reasoning for Domain-Specific Artifact Scenes (CD-VLR), a cross-modal interaction modeling method designed for complex cultural heritage environments. The core of the proposed approach lies in the design of a multimodal feature fusion module and an adaptive denoising mechanism, which together enable deep semantic alignment between images and texts while maintaining stable performance under noisy conditions. Experimental results demonstrate that the proposed method achieves an accuracy of 93.1% and an F1-score of 90.8% on the cross-modal image-text retrieval task using a self-constructed cultural heritage dataset, outperforming advanced baseline methods and exhibiting stronger robustness in noisy environments. This study provides an effective technical framework for cross-modal understanding of cultural heritage, advancing research on multimodal fusion methods and offering a feasible solution for digital archiving and intelligent interpretation of cultural heritage.