Large-scale dataset generation technology for oracle object detection

Oracle bone inscription object detection is an important part of oracle bone inscription digitization research. This work mainly relies on deep learning models to realize the recognition of position information and classification information in oracle bone inscription images. In order to avoid model overfitting, deep learning models need to rely on large-scale datasets. In the field of oracle bone inscription object detection, there are currently few large-scale data sets available for deep learning. Many research datasets rely on experts to manually annotate and organize, which makes oracle bone inscription object detection datasets face problems such as high cost, small data volume, low data quality, and poor balance between categories. This study proposes a dynamic two-stage Mosaic algorithm and oracle bone inscription large-scale dataset generation technology to solve the problems of limited number of mosaic images, insufficient image diversity and difference, large blank background, and missing information in the traditional Mosaic algorithm in processing oracle bone images. A complete dataset generation process is designed to realize the process-based and intelligent processing from oracle bone inscription single character images to dataset generation, which fundamentally solves the data dilemma in the field of oracle bone inscription object detection. Using the method in this study, a large-scale oracle bone inscription dataset with labeled position information and category information was generated. A total of 570000 oracle bone inscription images and 570000 corresponding annotation files were generated, including 416 oracle bone inscription categories, and the minimum category contained 516 oracle bone inscription characters. The dataset size and the number of samples in each category can be adjusted dynamically to avoid the problem of sample imbalance between categories. This research uses the YOLOv8 model to train the generated large-scale dataset. After 200 batches of training, the model precision reached 96.45%, the mAP50 value was 97.75%, and the mAP50-95 value was 96.96%. From the model training curve, the training process showed good stability and efficiency. The model training results show that the dataset generation technology in this paper can be applied to oracle bone inscription target detection research.