Ancient oracle bone inscription detection and translation based on depthwise-conv dual-cascade attention YOLOv12

Accurate detection and translation-oriented interpretation of oracle bone inscriptions are essential for the digital preservation and automated analysis of this invaluable cultural heritage. Existing deep learning approaches have primarily focused on detection challenges such as text adhesion, background noise, blurred strokes, and fragmented characters, while the connection between detected character regions and subsequent paleographic interpretation remains insufficiently explored. To address these challenges, this study proposes an enhanced YOLOv12-based framework for oracle bone inscription detection and translation-oriented analysis. The proposed framework incorporates three key innovations: a Depthwise-conv Dual-cascade Attention (DDA) module to enhance feature responses in critical stroke regions, a Ghost-Shuffle Convolution (GSC) module to improve feature fusion efficiency while reducing computational complexity, and a Scalable Intersection over Union (SIOU) loss function to improve localization precision. In addition, detected character regions are associated with corresponding modern Chinese character annotations, providing a preliminary bridge between detection and translation-oriented interpretation. Extensive experiments on large-scale oracle bone datasets demonstrate that the proposed detection framework achieves state-of-the-art performance, with 89.3% mAP@0.5 and 89.64% F1-score, while maintaining a practical model size of 48.5 MB. Comprehensive ablation studies, comparative experiments, training convergence analysis, and Grad-CAM visualizations further confirm the effectiveness of the proposed components in enhancing feature discrimination, feature fusion, and localization stability. This work provides a practical computational foundation for automated oracle bone inscription analysis and supports large-scale digital heritage conservation and archaeological research.