摘要
为了提高大数据中多模态信息的检索效果,提出一种基于深度神经网络的多模态信息检索算法。设计深度自编码器,将不同模态的数据投影到一个相同的广义子空间内;利用稀疏编码技术降低共同特征向量的维度,过滤冗余特征和噪声特征;通过去卷积操作和上采样操作对数据进行重建。基于公开模态识别数据集的实验结果表明,该算法能够有效地学习和泛化多模态数据,且在多模态检索实验中也表现出较好的性能。
In order to improve the retrieval effect of multimodal information for big data,this paper proposes a multimodal information retrieval algorithm based on deep neural network.A deep auto-encoder was designed to project multimodal data to a common general subspace;the sparse encoding technique was used to reduce the dimensions of the common feature vectors,so as to filter both redundancy features and noisy features;through the de-convolution and up-sampling operation,the data was reconstructed.The experimental results on public pattern recognition datasets show that the proposed algorithm can effectively learn and generalize multimodal data,and it performs better in multimodal retrieval experiments.
作者
李光宇
Li Guangyu(Information Office,Tianjin Academy of Fine Arts,Tianjin 300141,China)
出处
《计算机应用与软件》
北大核心
2022年第1期219-224,249,共7页
Computer Applications and Software
关键词
多模态学习
特征提取
压缩感知
稀疏编码
数据检索
深度学习
Multimodal learning
Feature extraction
Compressed sensing
Sparse encoding
Data retrieval
Deep learning
