Exploiting module evolution correlation relationship for fine-grained bird image classification with structural functional representation

Abstract

Fine-grained bird image classification (FBIC) targeting the classification of avian subspecies is facing challenges due to confusing background, partly occlusions and varied posture. In this paper, we proposed a novel module evolution correlation relationship modeling for FBIC task, which can learn structural functional representations among different functional feathers in a bird. Specially, the proposed EMECR model includes two modules, such as periodic topology mining (PTM), and multi-scale semantics alignment strategy (MSAS). The PTM module is proposed to reveal local periodic organizations with implicit functional expressions, and the MSAS is leveraged for better semantic modeling. In addition, a joint loss is designed to suppress the outliers and enforce semantic consistency. Furthermore, to better model the module evolution relationships between different functional representations and hierarchical context correlation, Mamba architecture is employed as the decoder with its linear computational complexity. Experiments on CUB-200–2011 and NABirds verify that our method can obtain robust results and significantly outperform the existing state-of-the-art FBIC methods. Extended experiments have been conducted on the Stanford Cars dataset to suggest the potential of generalizing our method on other fine-grained visual classification tasks.