Open AccessDynamic Multi-view Combination for Image Classification
Author(s) -
Chenghua Li,
Wanguo Wang,
Fengyuan Liu,
Zhenyu Guo
Publication year - 2020
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1631/1/012125
Subject(s) - discriminative model , computer science , artificial intelligence , image (mathematics) , pattern recognition (psychology) , contextual image classification , machine learning
Multi-view learning is widely used in image classification tasks to better explore the discriminative information of different views. However, existing multi-view methods commonly rely on some pre-defined assumptions or fail to fully take advantage of the combination commonality between individual images. This paper presents an efficient dynamic multi-view combination approach to dynamically combine the discriminative power of different views. Specially, we firstly utilize a group of pre-trained CNNs to extract different views of an image. Secondly, we apply a dynamic gating module to the image, which will generate a weight vector of these views to model the image-level information for the multi-view learning. Finally, the weight vector and the views are combined for the classification. Experimental results and analysis on CIFAR-10 and ImageNet show the effectiveness of the proposed dynamic multi-view combination method (DMVC) for visual classification.