MEMF: Multi-level-attention embedding and multi-layer-feature fusion model for person re-identification

作者：

Highlights：

• A multi-level-attention block cascading a spatial-level attention block with a channel-level attention block, is designed to obtain representative features.

• A multi-layer-feature fusion architecture is built, and a pooling block combination method is designed to obtain rich features.

• An eigenvalue difference orthogonality loss function is proposed to reduce the correlation between features and obtain discriminative features.

摘要

•A multi-level-attention embedding and multi-layer-feature fusion model is proposed for person re-identification. The multi-level-attention block is embedded in the feature extraction network, instead of being set as different branches.•A multi-level-attention block cascading a spatial-level attention block with a channel-level attention block, is designed to obtain representative features.•A multi-layer-feature fusion architecture is built, and a pooling block combination method is designed to obtain rich features.•An eigenvalue difference orthogonality loss function is proposed to reduce the correlation between features and obtain discriminative features.

论文关键词：Person re-identification,Feature expression,Convolutional neural network

论文评审过程：Received 20 February 2020, Revised 20 January 2021, Accepted 7 March 2021, Available online 11 March 2021, Version of Record 22 March 2021.

论文官网地址：https://doi.org/10.1016/j.patcog.2021.107937