Training-free block joint surface semantic recognition via topology-guided vision-language model for hull block assembly quality control

Dimensional accuracy control in hull block construction increasingly relies on 3D laser scanning for construction quality inspection and virtual assembly. In these tasks, the block joint surface is a critical region because it directly determines the geometric compatibility between adjacent blocks. Semantic segmentation of components in block joint surface is therefore essential for automated key point extraction, construction deviation evaluation, and virtual assembly. However, existing data-driven segmentation methods usually require task-specific annotated datasets and supervised training, which limits their generalization and rapid deployment in shipyard scenarios. To address this issue, this paper proposes a training-free semantic recognition framework for beam/shell binary classification of block joint surfaces. The proposed method exploits the quasi-planar contour characteristics of block joint surfaces and reformulates the 3D structural recognition task as 2D candidate mask extraction and topological relationship judgment. First, the joint surface design model is projected into a 2D binary contour image, and structure-aware adaptive cropping windows are generated to preserve small beam components within appropriate local fields of view. Then, the Segment Anything Model (SAM) is used to extract class-agnostic candidate masks. Finally, instead of directly querying professional component categories, this study proposes a prompting strategy based on topological relationships that abstracts beam components as branch-like structures attached to the main shell structure, and guides an open-source and locally deployable Vision-Language Model (VLM) to determine whether each candidate mask satisfies this topological definition. Experimental results demonstrate that the proposed method achieves accurate beam/shell recognition without task-specific training samples. For Qwen3 III, topology-based prompting improves the Mean PA and Mean mIoU from 0.5438 to 0.3840 to 0.9667 and 0.9087, respectively. In engineering experiments of ring units, the refined segmentation of design model achieves all four metrics above 0.99, while the semantic metrics of scanning point clouds after local registration exceed 0.98. These results, together with construction quality analysis and virtual assembly verification, indicate that the proposed framework, without relying on task-specific dataset collection and annotation, has strong engineering practicality and enables rapid deployment in shipbuilding applications.