Batch covariance neural network for image recognition

作者：

Highlights：

• Batch covariance layer is proposed to replace convolutional layer for improvement.

• BCL is regarded as a 3D covariance layer normalizing the kernels in batch size.

• BCL mitigates abnormal local features for discriminating and generating tasks.

• BCL can improve CNN's performance with thimbleful complexity increase.

摘要

•Batch covariance layer is proposed to replace convolutional layer for improvement.•BCL is regarded as a 3D covariance layer normalizing the kernels in batch size.•BCL mitigates abnormal local features for discriminating and generating tasks.•BCL can improve CNN's performance with thimbleful complexity increase.

论文关键词：CNN,BCovNN,Batch covariance,Illumination intensity,Feature interaction

论文评审过程：Received 5 October 2021, Revised 5 January 2022, Accepted 28 March 2022, Available online 8 April 2022, Version of Record 23 April 2022.

论文官网地址：https://doi.org/10.1016/j.imavis.2022.104446