Abstract
The planet is confronting a biodiversity crisis of historic proportions–species extinction rates are hundreds of times higher than natural levels, and ecosystem services continue to deteriorate. Efficient and accurate species monitoring technologies have become a core requirement for ecological conservation. Fine-grained visual recognition (FGVR) can automatically identify highly similar species (such as different bird subspecies, butterfly variants), providing a scalable and low-cost technological pathway for biodiversity monitoring. Although many methods, especially those based on transformers, have made a certain progress, they still struggle with complicated background interference, imbalanced attention allocation and limited cross-domain transferability. This paper proposes a Frequency-Aware Token-Filtered Transformer (FT-Former), featuring: (1) A Frequency Token Extraction module that harnesses spectral-domain features for maintaining critical discriminative details while compensating for information loss from spatial down-sampling operation, (2) A Hierarchical Token Filtering Mechanism that dynamically filters high-value tokens and intelligently distributes attention to balance feature preservation with computational efficiency, and (3) A Token-Level Cache Unit that stores the input-initialized tokens to prevent the loss of semantically important yet attention-weak discriminative features. Experimental results indicate that FT-Former achieves state-of-the-art performance across multiple benchmarks, with top1 accuracy of 93.2% on CUB-200-2011, 93.1% on NABirds (ranking first on both official leaderboards), and 89.2% on the challenging iNaturalist 2018 dataset.