Ancient character detection based on fine-grained density map

Ancient character detection is a crucial step for automatic recognition and digital library management. However, inscriptions on ancient artifacts vary in size and are sparsely distributed, challenging the detection of small characters with complex noise. This paper proposes a Fine-grained Density Map (FDM)-based method for ancient character detection. We design a semantic-based density map generation network to locate characters, crop the original image based on the density map, and fuse the detection results of cropped images with the original image. In addition, we propose a non-maximum suppression filtering strategy based on a density map (NMS-DM) to utilize the location information from the density map. Compared with the baseline, our method improves F1-score and recall by 2.16% and 6.10% respectively, while also improving precision and recall on small characters by 2.20% each. Our method can improve detection performance, especially for ancient scripts with complex background noise and numerous small characters.