Research on cascade detection of Chinese inscription text

Inscriptions serve as vital carriers of historical and cultural heritage. Automated inscription detection can reduce manual costs and improve conservation efficiency. However, existing approaches remains constrained by data scarcity, a lack of domain-specific architectures, and the absence of systematic detection pipelines. To address these limitations, we propose a cascaded detection method to support downstream tasks. Specifically, we first construct a structured inscription dataset encompassing diverse calligraphic styles. We then design a hierarchical line-to-character detection pipeline, optimizing the model architecture to accommodate the distinct textual characteristics of inscriptions alongside task-oriented data processing strategies. Extensive experiments on inscription images and related scenarios validate the effectiveness and feasibility of the proposed method. Integrating the enhanced model into the cascaded pipeline effectively mitigates text under-coverage, achieving a reduction in false positives of 47.7% and 49.5% across the two detection stages while preserving computational efficiency and demonstrating strong cross-domain generalization potential on ancient-book data.