MADC-ED: A Multimodal Approach for Dimensional Classification in Engineering Drawings

Semantic understanding of dimensional annotations in engineering drawings is crucial in fields such as industrial design, manufacturing, and automated quality control. However, existing methods still face limitations in handling the diversity of annotation styles and capturing the spatial context of annotations. To address these challenges, this paper proposes MADC-ED, a multimodal learning-based approach for semantic understanding of dimensional annotations. The method integrates image features, textual descriptions, arrow positions, and relative positional information to accurately interpret the semantic meaning of dimensional annotations. Experimental results on a BGA package three-view drawing dataset show that MADC- ED outperforms single-modality methods by over 11 % in accuracy. Furthermore, the method demonstrates good scalability and can be extended to semantic understanding tasks for other types of engineering drawings, highlighting its potential for broader application, especially in industrial contexts.