Deep variance network: An iterative, improved CNN framework for unbalanced training datasets

作者：

Highlights：

• We propose a novel deep variance network (DVN) by integrating subspaces with Bayesian network into CNN framework.

• We propose a hierarchical Bayesian model for unbalance learning of inner-class heterogeneity and inter-class homogeneity.

• We generate virtual samples to complete the unbalanced dataset in a top-down way from feature level to image level.

摘要

•We propose a novel deep variance network (DVN) by integrating subspaces with Bayesian network into CNN framework.•We propose a hierarchical Bayesian model for unbalance learning of inner-class heterogeneity and inter-class homogeneity.•We generate virtual samples to complete the unbalanced dataset in a top-down way from feature level to image level.

论文关键词：Deep variance network,Unbalanced training datasets,Convolutional neural network,Homogeneity,Heterogeneity

论文评审过程：Received 15 August 2017, Revised 18 March 2018, Accepted 27 March 2018, Available online 4 April 2018, Version of Record 14 April 2018.

论文官网地址：https://doi.org/10.1016/j.patcog.2018.03.035