Abstract
Han pictorial stones constitute an important component of China’s ancient cultural heritage, and their depictions of chariots, horses, and human figures provide valuable visual evidence for investigating travel activities in the Han dynasty. The identification and localization of chariot-and-horse elements provide a foundation for the retrieval and statistical analysis of relevant materials. However, current documentation practices rely largely on manual identification and transcription, limiting the efficiency of large-scale processing. In Han pictorial stone images, weak contours, textural interference, dense occlusion, and scale variation pose substantial challenges for reliable object detection. Accordingly, 539 images were collected and annotated with 3736 instances. Building on this dataset, an improved YOLO11n model was proposed, incorporating the RubbingDown multi-branch downsampling structure, C3k2GRM module, and task-adapted ASF multi-scale fusion pathway for information preservation, contextual representation, and cross-scale feature fusion. The proposed model achieved mAP@0.5, mAP@0.5:0.95, and recall of 88.0%, 54.5%, and 86.5%, respectively, representing improvements of 5.5, 2.8, and 9.9 percentage points, respectively, over the YOLO11n baseline. The model maintained 2.58 M parameters, while GFLOPs increased from 6.3 to 7.3. The detection outputs could support future workflows for documenting, retrieving, quantitatively analyzing, and manually verifying chariot-and-horse elements in Han pictorial stone images.