Pyramid convolution and multi-frequency spatial attention for fine-grained visual categorization

Abstract

Learning diverse detailed feature is crucial for fine-grained visual categorization (FGVC). However, most of existing methods for FGVC use the standard convolution for feature extraction which leads to the loss of many subtle but important features. Besides, in the existing attention models for FGVC, the features are aggregated by a simple global average pooling operation, which is unable to characterize complex feature information. To address these problems, we propose pyramid convolution and multi-frequency spatial attention (PCMFSA) for FGVC. To capture spatial context dependencies of different levels and avoid the loss of the subtle features, pyramid convolution, which uses convolution kernels with various sizes, is introduced in the backbone network. To reduce the influence of background noise and obtain diverse fine-grained feature information, we propose multi-frequency spatial attention (MFSA) module, in which two channel-wise grouping operations are designed and discrete cosine transform (DCT) in the frequency domain is adopted to enhance features of different spatial positions. Moreover, to mine more valuable regions, peak suppression is used to suppress high-response regions after MFSA. A large number of experimental results show that our method achieves state-of-the-art results on three popularly FGVC datasets.