A cross-granularity feature fusion method for fine-grained image recognition

A cross-granularity feature fusion method for fine-grained image recognition

연구 분야: Artificial Intelligence

학회: Applied Intelligence

Fine-grained image recognition is characterized by high interclass object similarities and large intraclass object variations. Many existing works focus on locating more discriminative parts, but it is difficult to extract multigranular features synchronously and fuse them to make joint decisions about various granular parts. To address these issues, this work proposes a novel cross-granularity feature fusion method. First, a multi-granularity feature generator is used to obtain various granularity features simultaneously for mid-level feature maps via its subgenerators. The subgenerators divide the feature maps into blocks to ensure the relative integrity of the local features, and randomly shuffle the divided blocks to increase the variance of the local regions. Then, a cross-granularity feature fusion strategy achieves the joint decision-making of multiple granularity features in fine-grained images. Therefore, the proposed method can extract various granularity features and promote the synergistic interaction of richer granularity features. The effectiveness of the method is verified through comprehensive experiments on three widely-used fine-grained object recognition benchmark datasets and a chip inner structure dataset. The experimental results show that the proposed method significantly outperforms the baseline and exhibits a comparable performance to that of the SOTA method. Source codes are available at https://github.com/ShanWuJ/CGFF