Open AccessDynamic Feature Combination by Agreement for Image Classification
Author(s) -
Chenghua Li,
Wanguo Wang,
Linzhi Liu,
Tian Liang
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 - 10.1088/1742-6596/1631/1/012151
Subject(s) - computer science , feature (linguistics) , weighting , image (mathematics) , focus (optics) , artificial intelligence , pattern recognition (psychology) , class (philosophy) , task (project management) , machine learning , contextual image classification , data mining , engineering , medicine , philosophy , linguistics , physics , systems engineering , optics , radiology
Image classification is a basic and important task in computer vision. Recently, various neural networks have been designed and proved to be very powerful models for image classification. It is natural for thinking of how to gather their strengths together, which refers to feature combination tasks. Traditional combination methods mainly focus on designing specific combination algorithms to achieve higher performance. However, few works consider how to utilize their agreement on a given target (for example, a specific class) to achieve better combinations. This paper presents a novel dynamic feature combination method (DFCA) for image classification problems based on the agreement of the individual features. DFCA promisingly takes the agreement of not only the commonalities, but also the individualities of different features by dynamically updating the weighting coefficients of given features using a routing module. Experiment and extensive analysis on CIFAR-10 prove the effectiveness and promising characteristics of the proposed method.