2026 Volume 7 Issue 3 Pages 768-784
This study investigates a method for constructing a feature space for retrieval and attribute recognition through contrastive learning using bridge inspection images and damage reports. Although Contrastive Language-Image Pre-training (CLIP) has demonstrated strong performance on general image-text pairs, its feature space is not necessarily well suited to specialized domains such as bridge inspection. To address this issue, we fine-tune a pretrained CLIP model using approximately 130,000 bridge inspection image–damage report pairs and construct a feature space specialized for bridge inspection. In the experiments, we evaluate retrieval performance and examine attribute recognition performance in terms of structural component type, damage type, condition rating, and maintenance action category. Experimental results show that the proposed method achieves better retrieval performance than the pretrained CLIP model and is also effective for recognizing attributes related to bridge inspection.