.png){kind=icon}
返回
Deep mutual information multi-view representation for visual recognition
.png){kind=icon}
PRE
AI
DOI：10.1007/s10489-022-03462-y.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
Multi-view representation is a crucial but challenging issue in visual recognition task. To address this issue, a deep mutual information multi-view representation method is proposed in this paper. Firstly, multi-view inputs are fed to the encoder module of the variational auto encoder architecture to extract multi-view latent layer features. Secondly, the correlation between local features and latent layer features of each view is calculated by maximizing the mutual information. Meanwhile, to obtain a robust multi-view representation, the multi-view canonical correlation analysis and the mutual information maximization methods are used to calculate the canonical correlation of different view mean vectors and the information correlation of different view distributions, respectively. Finally, the supervised loss is used to improve the discriminability of the middle feature layers. The proposed method can obtain a more robust hidden layer representations and operate multi-view scenes with more than two views. Experimental results demonstrate that the proposed method achieves better recognition accuracy than other compared methods among five publicly available datasets.
Keyword:
Multi-view representation
Mutual information
Multi-view canonical correlation analysis
Variational auto encoder
Neural network
期刊
IF:
3.5
论文数:
7.6K
被引数:
1.7W
机构
引用论文
.png){kind=icon}
查看更多内容