DPGT: An End-to-End Deformable Pairwise Graph Transformer for Dense P&ID Graph Reconstruction

Automatic digitization of Piping and Instrumentation Diagrams (P&IDs) is important for process automation and topology-based analysis in industrial plants. However, conventional modular approaches sequentially perform object detection, pipeline extraction, and relation construction, allowing errors from earlier stages to propagate to subsequent stages. This study proposes the Deformable Pairwise Graph Transformer (DPGT), an end-to-end framework that integrates deformable-attention-based object detection with exhaustive pairwise relation inference. DPGT structurally separates object prediction from relation prediction and assigns an independent learnable relation token to each object pair. In addition, Connector nodes are explicitly modeled to represent physical pipeline junctions and preserve detailed connectivity paths. The performance analysis in this study was conducted on the synthetic Dataset-P&ID benchmark. The configuration without Connector nodes achieved an Object mAP@0.5 of 0.8375 and an Average Edge Recall@300 of 0.9602, while the configuration with Connector nodes achieved 0.8378 and 0.9165, respectively. RelationFormer employs a single global relation token for relation modeling, whereas the proposed DPGT adopts an N × N object-pair structure in which the relation of each object pair is explicitly inferred. These results indicate that DPGT maintains stable object detection performance while reconstructing dense P&ID graphs with explicit physical connectivity and detailed topological information.