Research on segmentation and recognition of single characters of oracle bone script based on YOLOv5 and ResNeSt50

Oracle bone script, as the earliest mature writing system in China, provides critical insights into the origin of Chinese characters and ancient civilizations. Current recognition methods, however, struggle with insufficient feature discrimination and low accuracy for visually similar glyphs. To address these limitations, we propose a three-stage preprocessing pipeline: (1) binarization for contrast enhancement; (2) mean filtering for noise suppression; (3) fractional-order differential processing for edge sharpening. In this paper, we construct an oracle bone script character segmentation model based on the YOLOv5 algorithm, and add the convolutional attention module to the neck and head networks, achieving 98.5% precision and 97.2% mean average precision. For the recognition classification method, a coordinate attention mechanism module is introduced into the ResNeSt50 network. By weighting and fusing the features obtained in width and height, better visualization can be obtained.