Dynamic Concept-Aware Network for Few-Shot Learning

Jun Zhou,Qiujie Lv,Calvin Yu-Chian Chen
DOI: https://doi.org/10.1016/j.knosys.2022.110045
IF: 8.139
2022-01-01
Knowledge-Based Systems
Abstract:Few-shot learning (FSL) aims to recognize novel classes with extremely limited samples via adapting the prior knowledge learned from base classes. Most of existing methods for feature alignment in FSL consider the correspondence of semantic concepts between support and query images. However, encoding task-specific information for query features has not been sufficiently explored. Therefore, in this paper, we propose a dynamic concept-aware network (DCAN), which efficiently encodes task-specific structural concepts and adaptively dynamic alignment. Concretely, this is achieved by dynamic prototype task awareness (DPTA) and cross-correlation dynamic alignment (CoDA). The DPTA module assigns concept weights to each pixel position feature of support images so that dynamically generate prototypes to encode task-specific information with attention mechanism. The CoDA module first calculates the co-attention maps between image representations, and adaptively learns channel-dependent and spatial-dependent dynamic meta-filters based on inputs. Combining two dynamic modules to obtain final embeddings that generalize well to novel categories. Extensive ex-periments demonstrate that DCAN outperforms the state-of-the-art methods on four FSL classification benchmarks. Our code is available at https://github.com/zjgans/dcan.(c) 2022 Elsevier B.V. All rights reserved.
What problem does this paper attempt to address?