Few-Shot Incremental Learning With Context-Aware Spatial Enhancement for Image Recognition
Few-shot incremental learning (FSIL) refers to the ability of a model to learn new concepts from a limited number of labeled examples and gradually recognize novel categories with minimal supervision while retaining previously learned knowledge to prevent forgetting. To address the key challenges in...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11036738/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Few-shot incremental learning (FSIL) refers to the ability of a model to learn new concepts from a limited number of labeled examples and gradually recognize novel categories with minimal supervision while retaining previously learned knowledge to prevent forgetting. To address the key challenges in FSIL, this paper proposes a novel Context-Aware Spatial Enhancement (CASE) framework, which improves feature representations by jointly leveraging global and local spatial information. Specifically, the global spatial enhancement module captures long-range dependencies to enrich semantic context, while the local spatial enhancement module applies Gaussian filtering to refine fine-grained details and suppress background noise. Additionally, a multi-head interaction block is introduced to model intricate relationships between spatial regions, effectively bridging global and local perspectives for robust and context-aware feature learning. CASE also incorporates prior knowledge from base categories to enhance adaptability to novel classes. This comprehensive design not only reduces background interference and highlights salient features but also improves generalization and stability in few-shot learning scenarios. Extensive experiments on benchmark datasets validate the effectiveness of the proposed approach, demonstrating competitive performance along with superior scalability and generalization capability in FSIL tasks. |
|---|---|
| ISSN: | 2169-3536 |