A lightweight Transformer guided by features from multiple receptive fields for few-shot fine-grained image classification

Weichuan Zhang, Jiale Wang, Tao Lei, Zicheng Pan, Mohammad Aminul Islam, Yongsheng Gao, Changming Sun

Keywords: Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), Few-shot Learning, End-to-end training

Abstract: Convolutional neural networks (CNNs) and vision Transformers (ViTs) play key roles in few-shot fine-grained image classification (FSFGIC). One of the main challenges of FSFGIC is how to consistently learn high-quality feature representations from different very limited fine-grained datasets. CNNs struggle with long-range dependencies due to their inherent localized receptive fields, and ViTs might impair high-frequency information, e.g., local texture information. Furthermore, ViTs require a large number of training samples to infer feature properties such as translation invariance, locality, and the hierarchy of visual data, while FSFGIC's training samples are extremely limited. To address the problems mentioned, a new lightweight Transformer guided by features from multiple receptive fields (LT-FMRF) is proposed which has considered how to manage long-range dependencies and how to extract local features with multiple scales, global features, and fused features from input images for increasing inter-class differences and consistently obtaining high-quality feature representations from different types of limited training datasets. Furthermore, the proposed LT-FMRF can be easily embedded into a given few-shot episodic training mechanism for end-to-end training from scratch. Experimental results conducted on five widely used FSFGIC datasets consistently show significant improvements over twenty state-of-the-art end-to-end training-based methods.

Primary Area: transfer learning, meta learning, and lifelong 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.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 163