TransIFC: Invariant Cues-aware Feature Concentration Learning for Efficient Fine-grained Bird Image Classification

Fine-grained bird image classification (FBIC) is notonly meaningful for endangered bird observation and protectionbut also a prevalent task for image classification in multimediaprocessing and computer vision. However, FBIC suffers fromseveral challenges, such as bird molting, complex background, andarbitrary bird posture. To effectively tackle these challenges, wepresent a novel invariant cues-aware feature concentrationTransformer (TransIFC), which learns invariant and coreinformation in bird images. To this end, two novel modules areproposed to leverage the characteristics of bird images, namely,the hierarchy stage feature aggregation (HSFA) module and thefeature in feature abstraction (FFA) module. The HSFA moduleaggregates the multiscale information of bird images byconcatenating multilayer features. The FFA module extracts theinvariant cues of birds through feature selection based ondiscrimination scores. Transformer is employed as the backboneto reveal the long-dependent semantic relationships in birdimages. Moreover, abundant visualizations are provided to provethe interpretability of the HSFA and FFA modules in TransIFC.Comprehensive experiments demonstrate that TransIFC canachieve state-of-the-art performance on the CUB-200-2011 dataset(91.0%) and the NABirds dataset (90.9%). Finally, extendedexperiments have been conducted on the Stanford Cars dataset tosuggest the potential of generalizing our method on other finegrained visual classification tasks.