Common Feature Learning for Zero-shot Image Recognition

Shuang Li, Lichun WANG, Kai Xu, Jianjia Xin

Keywords: Zero-shot Image Recognition；Visual-semantic Relationship；Fine-grained Alignment；Semantic Vectors Generation；

Abstract:

The key issue of zero-shot image recognition (ZIR) is how to infer the relationship between visual space and semantic space from seen classes, and then effectively transfer the relationship to unseen classes. Recently, most methods have focused on how to use images and class semantic vectors or class names to learn the relationship between visual space and semantic space. The relationship established by these two methods is class-level and coarse-grained. The differences between images of the same class are ignored, which leads to insufficiently tight relationships and affects the accurate recognition of unseen classes.To tackle such problem, we propose Common Feature learning for Zero-shot Image Recognition (CF-ZIR) method to learn fine-grained visual semantic relationships at the image-level. Based on the inter class association information provided by class semantic vectors, guide the extraction of common visual features between classes to obtain image semantic vectors. Experiments on three widely used benchmark datasets show the effectiveness of the proposed approach.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 14222