MSDFTR:多阶段双分支融合的西夏文字识别方法  

MSDFTR:Multi-stage dual-branch fusion for recognizing Tangut characters

在线阅读下载全文

作  者:马金林 闫琦[1] 马自萍 MA Jin-lin;YAN Qi;MA Zi-ping(School of Computer Science and Engineering,North Minzu University,Yinchuan 750021,China;Key Laboratory for Intelligent Processing of Computer Images and Graphics of the State Ethnic Affairs Commission of PRC,North Minzu University,Yinchuan 750021,China;School of Mathematics and Information Science,North Minzu University,Yinchuan 750021,China)

机构地区:[1]北方民族大学计算机科学与工程学院,宁夏银川750021 [2]北方民族大学图像图形智能信息处理国家民委重点实验室,宁夏银川750021 [3]北方民族大学数学与信息科学学院,宁夏银川750021

出  处:《计算机工程与设计》2024年第11期3390-3396,共7页Computer Engineering and Design

基  金:北方民族大学中央高校基本科研业务费专项基金项目(2021KJCX09);宁夏自然科学基金项目(2023AAC03264、2022AAC03268);国家民委图像与智能信息处理创新团队开放课题基金项目(2022KF01)。

摘  要:针对因字形复杂和图片质量不高导致的西夏文字识别准确率不佳的问题,提出一种多阶段双分支西夏文字识别方法MSDFTR。提出一种关注通道特征的CSA注意力机制与关注空间特征的SDA注意力机制,采用CSA与SDA分别构建提取西夏文字通道特征和空间特征的逆残差瓶颈模块。使用多阶段特征提取方式分阶段捕捉图像中的有效特征,增强特征重用和特征表达能力。为增强模型鲁棒性与可解释性,基于通道和空间特征提出一种双分支网络结构。使用密集Transformer块深入融合多层特征。实验结果表明,MSDFTR在TCD-E数据集上的准确率达99.43%,比其它方法更高。To address the problem of poor recognition accuracy of Tangut characters due to the complexity of glyphs and low qua-lity of images,a multi-stage dual-branch Tangut character recognition method,MSDFTR,was proposed.CSA attention mechanism focusing on the channel features and SDA attention mechanism focusing on the spatial features were proposed,and the inverse residual bottleneck module for extracting the channel features and spatial features of the Tangut characters was constructed using CSA and SDA,respectively.The first step was to use multi-stage feature extraction to extract the channel features and spatial features.The effective features in the image were captured in stages using multi-stage feature extraction to enhance the feature reuse and feature expression capabilities.To enhance the model’s robustness and interpretability,a dual-branch network structure was proposed based on channel and spatial features.A dense Transformer block was used to deeply fuse multi-layer features.Experimental results show that MSDFTR achieves 99.43%accuracy on the TCD-E dataset,which is higher than that of other methods.

关 键 词:西夏文字识别 多阶段 特征融合 深度学习 逆残差块 通道特征 空间特征 

分 类 号:TP391.4[自动化与计算机技术—计算机应用技术]

 

参考文献:

正在载入数据...

 

二级参考文献:

正在载入数据...

 

耦合文献:

正在载入数据...

 

引证文献:

正在载入数据...

 

二级引证文献:

正在载入数据...

 

同被引文献:

正在载入数据...

 

相关期刊文献:

正在载入数据...

相关的主题
相关的作者对象
相关的机构对象