Task-aware Local Representation Mining Network for Few-shot Fine-grained Recognition

Lihan Lin, Jin Deng, Ying Huang, Yanyan Chen, Wei Luo

Abstract: Few-shot fine-grained learning aims at classifying samples into new unseen classes with only a handful of labeled samples available. The main challenge is the features of these classes often exhibit small inter-class differences or large intra-class differences.Currently, an effective method is to classify based on the discriminative Local Representations (LRs) mined from images. However, existing methods mostly consider only the information within the image itself when selecting LRs, overlooking the importance of considering the contextual semantic information between tasks to select the most discriminative LRs for each task. To address this issue, we propose the Task-aware Local Representations Mining Network (TRMN). In this work, we introduce a Meta-Filter Learner that generates a Meta-Filter based on the contextual semantic information between tasks to adaptively mine the most discriminative LRs for different tasks. Additionally, we incorporate a Meta-Attention Module that suppresses noise information in the images while enhancing focus on discriminative local features.Extensive experiments on benchmark datasets validate the superiority of our TRMN and demonstrate that our TRMN significantly outperforms most methods based on LRs.The source code is available at https://github.com/Accelerator-lh/TRMN.