Fine-grained feature collaborative constraints for multi-class character detection in Dongba manuscripts

Existing studies on Dongba manuscript character detection mainly focus on locating character regions rather than directly predicting categories. To support end-to-end multi-class detection, this paper constructs MDD814, a large-scale dataset containing 9000 sentence-level Dongba manuscript images, 57,563 annotated character instances, and 814 categories. To address high inter-class similarity, large intra-class variation, and severe long-tailed distribution, we propose FCC-Net, a Deformable-DETR-based detector with fine-grained feature collaborative constraints. FCC-Net introduces region-category prototype, query-category prototype, and query instance constraints to enhance fine-grained feature discrimination and stabilize category boundaries. Experiments on MDD814 show that FCC-Net reduces missed detections and misclassifications and achieves 71.84% AP, providing an effective technical basis for structured parsing and digital analysis of Dongba manuscripts.