Bi-focus cosine complement network for few-shot fine-grained image classification

Abstract

Few-shot fine-grained image classification presents significant challenges due to the need to discern subtle differences with few samples. And the image backgrounds often contain substantial redundant information. Existing methods utilize features containing redundant information directly after extraction to compute distances for classification, which diminishes the model’s object-level distinguishing capability. Additionally, these methods struggle to integrate discriminative and non-discriminative features within hierarchical architectures. To address this, we propose a Bi-focus Cosine Complement Network (BfCCN). In BfCCN, the Bi-focus Module (BFM) harnesses both spatial and channel dimension information to precisely localize targets within images by mitigating the impact of background noise on the model. Moreover, the Cosine Complement Network (CCN) leverages multi-level and multi-type features to capture the fine-grained details typically present in shallow layers but absent in deeper layers, thereby enhancing the model’s ability to detect subtle distinctions by utilizing complementary cosine distance for classification. Experiments on three benchmark fine-grained datasets demonstrate the efficacy of our method. Codes will be available at https://github.com/naivejph/BFCCN.git.