科技创新与应用2025，Vol.15Issue(11)：1-5,5.DOI:10.19981/j.CN23-1581/G3.2025.11.001
深度学习驱动的甲骨文拓片单字分割与识别
摘要
Abstract
In the field of protecting and inheriting Chinese cultural heritage,the digital processing of Oracle Bone Script(OBS)holds significant importance.This paper aims to achieve precise segmentation and recognition of OBS images through advanced image processing techniques.Initially,the original rubbing images are gray-scaled and pre-processed using the Otsu algorithm,morphological operations,and Gaussian blurring.Subsequently,feature extraction methods such as Histogram of Oriented Gradients(HOG),Scale-Invariant Feature Transform(SIFT),and Local Binary Patterns(LBP)are employed,in conjunction with the Random Forest algorithm,to construct a pre-processing model for OBS images.Then,a U-Net-based image segmentation model is designed and trained,with its performance comprehensively evaluated using multi-dimensional assessment indicators such as binary cross-entropy loss function and Dice coefficient.The model achieved an accuracy of 94.25%,demonstrating high segmentation precision.Finally,a single character recognition model based on the Feature Pyramid Network(FPN)is established and evaluated using metrics such as mean Average Precision(mAP),accuracy,recall,and F1 score.The test results show that the model's mAP reached 81.56%,highlighting its outstanding performance in OBS recognition and providing robust technical support for the digital preservation and research of Oracle Bone Script.关键词
甲骨文/图像分割/深度学习/特征提取/U-Net/文字识别/FPN目标检测Key words
Oracle Bone Script/image segmentation/deep learning/feature extraction/U-Net/single characterrecognition/FPN object detection分类
社会科学引用本文复制引用
邹阳,蒋志辉,张燕炜,许宏飞..深度学习驱动的甲骨文拓片单字分割与识别[J].科技创新与应用,2025,15(11):1-5,5.基金项目
2023年永州市指导性科技计划项目(2023YZ002) （2023YZ002）
湖南科技学院科学研究项目(23XKYZZ05) （23XKYZZ05）
湖南科技学院2024年校级大学生创新创业训练计划项目(无编号) （无编号）
