北京建筑大学学报2025，Vol.41Issue(3)：78-87,10.DOI:10.19740/j.2096-9872.2025.03.09
动态平衡融合的跨语言跨模态建筑信息检索
Dynamic Equilibrium Fusion for Cross-Lingual Cross-Modal Building Information Retrieval
摘要
Abstract
Building Information Retrieval(BIR)is designed to efficiently acquire corresponding architectural data through building images or textual descriptions.Existing methods encounter challenges,including difficulties in feature fusion and semantic misalignment,especially with multilingual text and complex architectural imagery.To address these issues,the Balanced Fusion Feature Network(BFFN)is proposed,which dynamically models multimodal fusion equilibria and adaptively captures deep feature correlations to encode intra-and inter-modal features across multiple levels.To enhance cross-lingual semantic understanding,knowledge distillation is introduced to extract shared knowledge from cross-lingual models,optimizing text encoder performance while reducing model complexity.Additionally,a dual-stream contrastive learning framework is proposed to strengthen the representation of positive samples and fused features,enabling fine-grained cross-modal semantic interactions and improving semantic alignment between multilingual text and architectural images.Experimental results demonstrate that the proposed model outperforms existing baselines on a cross-lingual architectural scene dataset while maintaining strong performance in both image and text retrieval tasks.关键词
建筑信息检索/对比学习/特征融合/知识蒸馏Key words
building information retrieval/contrastive learning/feature fusion/knowledge distillation分类
建筑与水利引用本文复制引用
张蕾,唐明亮,丁鑫,杨成伟..动态平衡融合的跨语言跨模态建筑信息检索[J].北京建筑大学学报,2025,41(3):78-87,10.基金项目
国家重点研发计划项目(2022YFB3305602) （2022YFB3305602）
教育部人文社会科学规划基金项目(22YJA630111,22YJAZH110) （22YJA630111,22YJAZH110）
北京建筑大学研究生创新项目(PG2024088) （PG2024088）
山东省重点研发计划项目(2021SFGC0102). （2021SFGC0102）
