Local Mutual Metric Network for Few-Shot Image Classification
| Autor: | Huaxiong Li, Chunlin Chen, Yaohui Li, Haoxing Chen |
|---|---|
| Rok vydání: | 2021 |
| Předmět: | |
| Zdroj: | Pattern Recognition and Computer Vision ISBN: 9783030880033 PRCV (1) |
| DOI: | 10.1007/978-3-030-88004-0_36 |
| Popis: | Few-shot image classification aims to recognize unseen categories with only a few labeled training samples. Recent metric-based approaches tend to represent each sample with a high-level semantic representation and make decisions according to the similarities between the query sample and support categories. However, high-level concepts are identified to be poor at generalizing to novel concepts that differ from previous seen concepts due to domain shifts. Moreover, most existing methods conduct one-way instance-level metric without involving more discriminative local relations. In this paper, we propose a Local Mutual Metric Network (LM2N), which combines low-level structural representations with high-level semantic representations by unifying all abstraction levels of the embedding network to achieve a balance between discrimination and generalization ability. We also propose a novel local mutual metric strategy to collect and reweight local relations in a bidirectional manner. Extensive experiments on five benchmark datasets (i.e. miniImageNet, tieredImageNet and three fine-grained datasets) show the superiority of our proposed method. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|