A fine-grained bidirectional multi-scale network for extreme-aspect-ratio bamboo slip character detection
Year: 2026
Type: article
Abstract: Automated bamboo slip character detection supports cultural heritage digitization, but it remains difficult because excavated manuscripts contain dense, slender, fragmented, and degraded characters. Faded ink, wood grain, cracks, and stains often resemble genuine strokes, causing missed detections, false positives, and unstable boundary localization. To address these challenges, we propose a Fine-Grained Bidirectional Multi-Scale YOLO (FBMS-YOLO) for extreme-aspect-ratio bamboo slip character detection. The framework integrates three task-oriented designs: an Attention-enhanced Refinement and Context Network (ARCNet) for structural representation of slender strokes, a Fine-grained Bidirectional Feature Pyramid Network (FB-FPN) for preserving high-resolution spatial details through adaptive cross-scale fusion, and a Dynamic Focusing and Boundary-Penalty Intersection over Union loss (DFB-IoU) for robust localization under ambiguous boundaries. Experiments on the DeepJiandu dataset show that FBMS-YOLO improves the YOLOv11-n baseline from 92.8 to 94.4% in mAP $$_{50}$$ and from 55.0 to 56.8% in mAP $$_{50:95}$$ . Additional component analysis, ablation studies, and cross-dataset evaluation on oracle bone inscription images demonstrate the effectiveness and transferability of the proposed framework. These results indicate that FBMS-YOLO provides an accuracy-oriented solution for dense character localization in degraded historical manuscripts.
Source: Discover Artificial Intelligence
Authors: Feng Gao, Chengjie Bao, Mengjie Kong, Bingquan Chen
Institutions: Anyang Normal University, Jishou University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,590
OpenAlex ID: W7214524324