We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
Dual-domain reciprocal learning design for few-shot image classification.
- Authors
Liu, Qifan; Chen, Yaozong; Cao, Wenming
- Abstract
Few-shot learning is challenging in computer vision tasks, which aims to learn novel visual concepts from few labeled samples. Metric-based learning methods are widely used in few-shot learning due to their simplicity and effectiveness. However, comparing the similarity of support samples and query samples in a single metric space appears to be biased. In this work, we design a dual-domain reciprocal metric network (DRM-Net) structure for few-shot classification task which establishes a commutative learning relationship in two feature distributions from different metric domains. Specifically, our reciprocal metric network contains two metric domains, which employ graph neural network (GNN) and geometric algebra graph neural network (GA-GNN) as two metric functions to comprehensively measure the similarity between samples. This structure can help reduce the prediction bias by a single measure. We also construct the reciprocal learning loss between the metric feature distributions from the two branches to promote each other to improve the performance of the overall model. Our extensive experimental results demonstrate that the proposed reciprocal metric learning outperforms existing state-of-the-art few-shot learning methods on various benchmark datasets.
- Subjects
IMAGE recognition (Computer vision); VISUAL learning; METRIC spaces; COMPUTER vision; PETRI nets; DIGITAL rights management
- Publication
Neural Computing & Applications, 2023, Vol 35, Issue 14, p10649
- ISSN
0941-0643
- Publication type
Article
- DOI
10.1007/s00521-023-08255-z