DSC: Containment-Aware Instance Consolidation for Ultra-High-Resolution Railway Signal-Layout Image Analysis

High-resolution engineering drawings are commonly analyzed in overlapping tiles, which can produce several predictions showing different extents of the same object. Such observations may have low intersection-over-union despite substantial foreground containment. We propose document-space consolidation (DSC), which projects local instance predictions into drawing coordinates and applies class-wise box non-maximum suppression followed by confidence-prioritized mask intersection-over-minimum suppression. This two-stage procedure removes redundant observations while preserving the geometry of retained masks. On ten held-out railway signal-layout drawings containing 1755 annotated instances, the validation-selected configuration achieves precision 0.8994, recall 0.9470, and F1 0.9226. With identical candidates and fixed box-NMS/containment thresholds of 0.50/0.60, DSC reduces false positives from 288 to 186 and improves F1 from 0.8989 to 0.9226 relative to box-based containment filtering. Three-seed local-recognition ablations, a joint threshold analysis, and quantitative error and resource measurements characterize the evaluated pipeline. Optical character recognition and railway-specific rules are assessed separately as application refinement. These results support containment-aware consolidation for producing consistent instance inventories from overlapping tiles in the studied engineering drawings.