Recognition of Components in Taoist Charm Images

Taoist charms are used in Chinese culture for communication with spirits, offering healing, exorcism, and supernatural abilities. These charms, often crafted by Taoist priests, feature a variety of characters and symbols, differing across sects, which complicates manual recognition and study. This paper focuses on automating the recognition and retrieval of charm components to ease research efforts. By manually annotating 998 charm images into 107 component categories, we addressed data scarcity and imbalance through data augmentation, expanding the dataset significantly. We refined data labeling accuracy and minimized distortions in synthetic data, employing deep learning techniques for object detection and segmentation. After testing different models, we optimized our approach by pre-training the model’s backbone network on a large Chinese text dataset, enhancing the model’s ability to recognize charm components. Achieving a bounding box mAP of 0.764 and a mask mAP of 0.726.