YOLO-ACD: Robust Lightweight Detector for Ancient Character Detection

Ancient character detection serves an important role in paleographic studies and archaeological work. Unlike modern characters or text, ancient character renders common text detection algorithms inapplicable due to its complex arrangement patterns and highly variable backgrounds. Current methods commonly adopt general object detection models, which often struggle with the unique challenges of ancient characters. In this paper, we propose a lightweight ancient character detection method based on YOLO detector called YOLO-ACD. Starting from an efficient YOLO structure, to address the defects and adhesion due to coarse scanning or rubbing of the characters, we propose a novel high-frequency channel attention (HFCA) module to guide model for precise saliency extraction of strokes. To eliminate the interference of noisy backgrounds on the character area, we integrate a context-aware spatial attention (CASA) module. Furthermore, in order to overcome the limitations posed by limited training samples, we propose the incorporation of spectral clustering results as extra confidence scores, with the objective of enhancing the recall rate of the predictions. Extensive experiments on two ancient character datasets, including Chinese oracle bone script and Egyptian hieroglyphs, demonstrate that our method outperforms both mainstream and similar approaches. Compared to baseline model, the mAP is improved by 9.8-12.5%. The code of the proposed method will be available at GitHub.