[1]李子含,屈乐达,刘思源.基于FM-MobileViT网络的拓片甲骨文字识别[J].计算机技术与发展,2025,(05):23-28.[doi:10.20165/j.cnki.ISSN1673-629X.2025.0020]
　LI Zi-han,QU Le-da,LIU Si-yuan.Rubbing Oracle Bone Character Recognition Based on FM-MobileViT[J].,2025,(05):23-28.[doi:10.20165/j.cnki.ISSN1673-629X.2025.0020]
点击复制

基于FM-MobileViT网络的拓片甲骨文字识别()

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
期数:
2025年05期
页码:
23-28
栏目:
媒体计算
出版日期:
2025-05-10

文章信息/Info

Title:
Rubbing Oracle Bone Character Recognition Based on FM-MobileViT
文章编号:
1673-629X(2025)05-0023-06
作者:
李子含; 屈乐达; 刘思源
沈阳理工大学 信息科学与工程学院,辽宁 沈阳 110158
Author(s):
LI Zi-han; QU Le-da; LIU Si-yuan
School of Information Science and Engineering,Shenyang Ligong University,Shenyang 110158,China
关键词:
甲骨文字拓片; MobileViT; 文字识别; 深度学习; 融合模块; 注意力机制
Keywords:
rubbings of oracle bone character; MobileViT; character recognition; deep learning; fusion module; attention mechanism
分类号:
TP391
DOI:
10.20165/j.cnki.ISSN1673-629X.2025.0020
摘要:
甲骨文是汉字的源头和中华优秀传统文化的根脉。 由于甲骨文拓片为人工雕刻且深埋地下,存在样本分布不均、噪声严重等问题,导致识别精度不高。 针对以上问题,提出一种基于 FM - MobileViT ( Fusion and Attention Mechanism MobileViT)网络的甲骨文字识别方法。 首先,对数据集中图像进行锐化预处理操作,使目标边缘更清晰明显;并对数据数量过少的字符类别对应图像采用随机旋转、随机错切等方式进行数据增强,提升了数据集的质量,丰富了样本数据。 其次,设计融合模块,构建跳转连接结构,将深浅层特征融合,使提取到的特征图能够融合浅层特征和语义特征;并在融合模块中引入 CBAM 注意力机制,使融合操作更有指向性、目的性,增强模型特征提取的能力。 通过消融实验和对比实验表明,FM-MobileViT 模型识别准确率达到 92. 3% ,比 MobileViT 提升了 1. 7 百分点,同时 FPS 达到 30 107。 相比于同类型的网络结构,FM-MobileViT 不仅有更高的准确率,而且取得了精度与速度的权衡。
Abstract:
Oracle bone inscriptions are the source of Chinese characters and the root of Chinese excellent traditional culture. Because oracle rubbings are artificially carved and buried deep underground,they have some problems such as uneven sample distribution and serious noise,resulting in low recognition accuracy. To solve the above problems, we propose an oracle bone character recognition method based on FM-MobileViT(Fusion and Attention Mechanism MobileViT) network. Firstly,the image in the dataset is sharpened and preprocessed to make the target edge more clear and obvious. The image corresponding to the character category with too little data is enhanced by random rotation and random miscut,which improves the quality of the data set and enriches the sample data. Secondly,the fusion module is designed,the jump connection structure is built,and the deep and shallow features are fused,so that the extracted feature map can integrate the shallow features and semantic features. The CBAM attention mechanism is introduced into the fusion module to make the fusion operation more directional and purposeful, and enhance the ability of feature extraction. Ablation and comparison experiments show that the recognition accuracy of the FM-MobileViT model proposed reaches 92. 3% ,1. 7 percentage points higher than that of MobileViT,and the FPS reaches 30 107. Compared with the same type of network structure,FM-MobileViT not only has a higher accuracy,but also achieves a trade-off between accuracy and speed.

相似文献/References:

[1]曹 欢,方 睿.基于深度学习的芒果病虫害分类识别[J].计算机技术与发展,2023,33(10):115.[doi:10. 3969 / j. issn. 1673-629X. 2023. 10. 018]
　CAO Huan,FANG Rui.Classification and Identification of Mango Diseases and Pests Based on Deep Learning[J].,2023,33(05):115.[doi:10. 3969 / j. issn. 1673-629X. 2023. 10. 018]
[2]王召龙,张洁.基于改进的YOLOv8轻量级火灾检测算法研究[J].计算机技术与发展,2024,34(10):61.[doi:10.20165/j.cnki.ISSN1673-629X.2024.0206]
　WANG Zhao-long,ZHANG Jie.Research on Lightweight Fire Detection Algorithm Based on Improved YOLOv8[J].,2024,34(05):61.[doi:10.20165/j.cnki.ISSN1673-629X.2024.0206]

更新日期/Last Update: 2025-05-10