Dongba Script Recognition Network Based on Global and Local Features
This paper presents a deep-learning method for recognizing handwritten Dongba (Naxi) pictographic characters, one of the world's few living pictographic scripts. Building on a Swin Transformer backbone, the authors add a Contextual Transformer (COT) module to strengthen local stroke-level attention, an Explicit Visual Center (EVC) module (a lightweight MLP paired with a Learnable Visual Center) to capture global structural information, and a Feature Pyramid Network (FPN) to fuse multi-scale features. The model is evaluated on DB1404, a large handwritten Dongba character dataset split into 'ordinary' (1,352 classes, 302,245 images) and 'highly similar' (46 classes, 8,053 images) subsets. Ablation experiments show each module improves accuracy individually, with all three combined reaching 98.41% and 98.37% Top-1 accuracy on the two subsets respectively. The proposed method also outperforms EfficientNet, MobileViT, and ConvNeXt baselines, reaching up to 99.54% Top-5 accuracy, demonstrating improved recognition of structurally complex and visually similar Dongba characters.
This page publishes the bibliographic record and the project's summary. The scanned article text belongs to its publisher and is shown to readers under a research-use policy; it is not reproduced in this record.