FMA-Net: Fuzzy Mutual Attention Networks for Fine-Grained Image Recognition

초록

Fine-Grained Image Recognition (FGIR) remains challenging due to subtle inter-class differences, complex back grounds, and scale variations. To address these challenges, we propose the Fuzzy Mutual Attention Networks (FMA-Net), integrating a Separable Convolution (SC) branch and a Self Attention (SA) branch via fuzzy logic-based mutual attention. Specifically, we introduce two key modules: the Fuzzy Attention Guidance (FAG) module for semantic-to-local guidance, and the Fuzzy Background Masking (FBM) module for local-to-semantic background suppression. The FAG module uses global semantic information from the SA branch to guide SC features toward dis criminative regions, enhancing localization accuracy. Conversely, the FBM module generates fuzzy masks from SC features to sup press irrelevant backgrounds in the SA branch, thus significantly improving global semantic purity. Unlike existing methods, these modules leverage fuzzy logic inference, enhancing interpretability and explicitly modeling cross-branch semantic interactions. Extensive experiments on benchmark datasets (CUB-200-2011, Stanford Dogs, Oxford Flowers, and Food-101) demonstrate that FMA-Net achieves competitive accuracy, outperforming existing methods by up to 2.5%. Comprehensive ablation studies confirm the effectiveness and interpretability of the proposed FMA-Net.

키워드