The efficient utilization of 2D-CAD drawing assets accumulated in the manufacturing industry is crucial for cost reduction and the inheritance of design knowledge. However, existing retrieval systems face two significant challenges: the prohibitive la...
The efficient utilization of 2D-CAD drawing assets accumulated in the manufacturing industry is crucial for cost reduction and the inheritance of design knowledge. However, existing retrieval systems face two significant challenges: the prohibitive labor costs re- quired for manual attribute registration, and the limitations in search accuracy due to reliance solely on geometric shapes, which fails to capture semantic nuances. Further- more, unlike general imagery, the interpretation of technical drawings requires advanced engineering knowledge. Consequently, acquiring large-scale labeled datasets, which are indispensable for deep learning applications, poses a significant challenge specific to this domain. To address these issues, this study proposes a novel drawing retrieval system that leverages deep learning to automate the extraction of both shape and attribute in- formation. The proposed system features an automated registration pipeline utilizing YOLO11 for region detection, Qwen2-VL for structuring title block attributes such as part names and materials, and Yomitoku for digitizing annotation information, thereby eliminating the need for manual data entry. Additionally, this study introduces a feature extraction model trained via con- trastive learning based on the CLIP framework. By utilizing part name text automati- cally extracted by a VLM as pseudo-labels for drawing images, the model learns semantic visual representations from a large-scale dataset of 20,000 unlabeled drawings without requiring manual annotation. In evaluation experiments using actual manufacturing drawings, the proposed method achieved a Recall@5 of 91.1%, significantly outperforming the baseline model trained on manually labeled data, which recorded 80.7%. Furthermore, the method achieved a Re- call@10 of 96.4%, demonstrating that the proposed approach effectively integrates visual and semantic information to realize high-precision retrieval while dramatically reducing implementation costs.