Margin embedding net for robust margin collaborative representation-based classification

doi:10.1016/j.patcog.2022.108991

Return

Margin embedding net for robust margin collaborative representation-based classification

2023-01-01

PRE

Zhichao Zheng

Huaijiang Sun *

Ying Zhou

DOI：10.1016/j.patcog.2022.108991

Original

Original request for help

Save

Abstract

En 中文

Collaborative Representation-based Classification method (CRC) shows great potential in classification task. However, redundancies in both features and samples limit the application of CRC seriously. The existing works only solve one of them and ignore the other, which leads to performance degradation. To address this problem, we explore collaborative representation mechanism and propose a classification method termed Robust Margin Collaborative Representation-based Classification (RMCRC) which uses a few but more representative robust marginal samples to eliminate redundancy between samples. As the performance of RMCRC is related to robust marginal samples and class separability assumption closely, we further propose a feature extraction method termed Margin Embedding Net (MEN) for RMCRC. In MEN, virtual samples are generated by a generative model to enhance effectiveness of robust marginal samples and generalizability of RMCRC. Then, an embedding network with triplet loss is used to elimi-nate the redundancy in features and ensure the assumption is satisfied. Specifically, we construct triplet according to the collaborative representation. Hence, MEN fits RMCRC very well. Extensive experimental results validate effectiveness of proposed method.(c) 2022 Elsevier Ltd. All rights reserved.

Keywords:

Collaborative representation

Feature extraction

Marginal sample

Image classification