Hierarchical Fine-Grained Visual Classification Leveraging Consistent Hierarchical Knowledge

Hierarchical fine-grained visual classification assigns multi-granularity labels to each object, forming a tree hierarchy. However, how to minimize the impact of coarse-grained classification errors on fine-grained classification and achieve high consistency remains challenging. Considering the human ability to progress from understanding generalized concepts to recognizing subtle differences between categories, the proposed novel hierarchy-aware conditional supervised learning method encodes such dependencies within its learned structure. The validity masks based on label hierarchy are designed to control the influence of coarse-grained classifications on fine-grained classifications. In this paper, the graph representation learning is explored to better utilize label hierarchy, integrating hierarchical structural information into the feature representation framework. Experiments on three standard fine-grained visual classification benchmark datasets demonstrate the effectiveness of the proposed method, significantly improving the consistency of hierarchical predictions while enhancing the model's understanding of label hierarchy compared with the state-of-the-art methods.