文献类型：期刊文献

中文题名：基于拓扑数据分析与卷积神经网络的特征融合方法

英文题名：A feature fusion method based on topological data analysis and convolutional neural network

作者：杨含[1];秦广军[1];刘子源[2];胡永庆[1];刘光南[1];戴庆龙[1]

第一作者：杨含

机构：[1]北京联合大学智慧城市学院,北京100101;[2]国家电力投资集团数字科技有限公司,北京102209

第一机构：北京联合大学智慧城市学院

年份：2025

卷号：42

期号：5

起止页码：624-630

中文期刊名：深圳大学学报(理工版)

外文期刊名：Journal of Shenzhen University(Science and Engineering)

收录：;北大核心:【北大核心2023】；

基金：国家自然科学基金资助项目(62172045);北京联合大学校级科研资助项目(ZK10202403)。

语种：中文

中文关键词：人工智能;模式识别;计算机神经网络;拓扑数据分析;卷积神经网络;注意力机制;计算机图象处理

外文关键词：artificial intelligence;pattern recognition;computer neural networks;topological data analysis;convolu-tional neural network;attention mechanism;computer image processing

摘要：针对卷积神经网络(convolutional neural networks,CNN)难以捕获和利用复杂高维数据的多维结构信息,限制了其特征学习能力的问题,提出一种融合了拓扑数据分析(topological data analysis,TDA)与CNN的特征融合方法——TDA-CNN.该方法将CNN捕获的数值分布特征与TDA提取的拓扑结构特征相融合,CNN通道负责提取数值分布特征,TDA通道专注于提取拓扑结构特征,然后,将这两类特征融合形成组合特征表示,并利用注意力机制自适应地学习每种特征的重要性权重,为后续全连接网络提供更全面的决策依据.在Intel Image、Gender Images和Chinese Calligraphy Styles by Calligraphers等数据集上的实验表明,TDA-CNN在改进特征聚类与识别关键特征方面表现出色,分别将基线模型VGG16、EfficientNet V2和DenseNet121的性能提升了21.89%、22.66%和8.26%,有效增强了模型的判别能力.
A feature fusion method,TDA-CNN,is proposed to address the problem of convolutional neural networks(CNNs)being difficult to capture and utilize the multidimensional structural information of complex high-dimensional data,which limits their feature learning ability.The numerical distribution features captured by CNNs with the topological structure features extracted by TDA,and the CNN channel is responsible for extracting numerical distribution features,while the TDA channel focuses on extracting topological structure features.The numerical distribution features and topological structure features are then fused to form a combined feature representation.An attention mechanism is introduced to adaptively learn the importance weight of each feature,thus providing a more robust foundation for decision-making in the fully connected layers.Experiments on datasets such as Intel Image,Gender Images,and Chinese Calligraphy Styles by Calligraphers show that TDA-CNN performs well in improving feature clustering and identifying key features,improving the performance of baseline models VGG16,EfficientNet V2,and DenseNet121 by 21.89%,22.66%,and 8.26%,effectively enhancing the discriminative ability of the model.