Structure-Preserving and Redundancy-Free Features Refinement for Generalized Zero-Shot Learning

Jian Ni, Yong LiaoICASSP 2023 · OpenReview

Abstract

Generalized zero-shot learning (GZSL) aims to recognize images from seen and unseen classes. Most models achieve competitive performance but still suffer from two problems: (1) Topological structure neglection; (2) Redundant information interference. In this paper, we propose a Structure-preserving and Redundancy-free Features Refinement model (referred to as SP-RFFR) to address these problems correspondingly in two modules: (1) Structure-preserving, to explicitly incorporate the topological structure into the learning of the latent space and the generator; (2) Redundancy-free features refinement, to remove the redundant information from the visual features and learn class- and semantically-relevant representations. To the best of our knowledge, this is the first work that incorporates topological structure preserving and redundancy-free features refinement into a unified framework for GZSL. Extensive experiments show that SP-RFFR outperforms the state-of-the-art methods on four benchmarks.
BibTeX
@inproceedings{ni2023structure,
  author = {Jian Ni and Yong Liao},
  title = {{Structure-Preserving and Redundancy-Free Features Refinement for Generalized Zero-Shot Learning}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023},
  url = {https://openreview.net/forum?id=J4G0XZYY3y},
}

Loading related content...

Figure