Edge-guided structural conditioning for Gong-Che symbol localization and page-level counting in aged documents

Abstract Reliable symbol measurement from aged Gong-Che documents is difficult because yellowing, faded ink,and broken strokes degrade local appearance before they fully destroy notation structure. We present StructGuide-YOLO, a paired RGB-edge detector designed for symbol localization and page-level counting on degraded manuscript pages. The method treats RGB as the primary semantic stream and uses a deterministically derived edge view as hierarchical structural guidance at three backbone scales. We also introduce GongChePair, a paired RGB-edge benchmark built from digital Gong-Che pages and scanned physical scores. The original collection contains 1,516 unique pages; after training-only offline augmentation, the final experimental protocol contains 2,164 training samples, 308 validation samples, and 155 test samples, giving 2,627 page-level paired samples. On GongChePair, StructGuide-YOLO reaches 0.865 mAP@0.5 and 0.438 mAP@[0.5:0.95], improving on YOLO11m by 5.4% and 7.9%, respectively. At the default confidence threshold of 0.25, it also reduces page-level count error to 9.70 MAE and 15.40 RMSE, compared with 11.10 and 18.50 for YOLO11m. Matched RGB-only, edge-only, simple concatenation, direct fusion, component-ablation, edge-sensitivity, degradation, repeatability, and source/book-family holdout experiments further show that the gain is associated with structural conditioning rather than only model capacity or a favorable preprocessing setting. These results support a measurement-scoped conclusion: deterministic structural guidance can improve Gong-Che symbol localization and page-level counting under the reported benchmark protocols, while end-to-end transcription and complete music-score understanding remain outside the evaluated scope