MAGNET: Multi-level feature guidance network for few-shot fine-grained image classification

Few-shot fine-grained image classification aims to distinguish highly similar categories with limited labeled samples. However, existing methods face three limitations. First, they fail to effectively feature according to the characteristics of each layer, overlooking the fine-grained structures in low-level features. Second, they handle high-level features simplistically, lacking the ability to share knowledge across multiple tasks, and they struggle with background redundancy in mid-level features, leading to overfitting. To this end, we propose a Multi-Level Feature Guidance Network (MAGNET), which integrates three core modules. The Primary Information Enhancement Module enhances low-level features by capturing fine-grained structural information and reinforcing them with high-level features. The Wavelet Attention Knowledge Guidance module applies wavelet transform for frequency-domain analysis of high-level features, while a multi-task-related knowledge transfer mechanism improves the model’s ability to share knowledge across tasks, enhancing generalization to new categories. The Background Filtering Module reduces background redundancy in mid-level features using high-level semantic information, mitigating overfitting. Extensive experiments on three benchmark datasets demonstrate that MAGNET outperforms existing methods. The source code is available at https://github.com/naivejph/MAGNET.git.