查看文章

Multi-granularity visual classification is a challenging task derived from traditional image recognition. Previous methods commonly use the features from the final convolutional layer to perform multi-granularity visual classification. However, the features required for different granularity label classification are not consistent. The finer the granularity of the label, the more detailed the features are needed. So, the key to multi-granularity visual classification is to extract effective features for different granularity levels. Generally, the high-frequency parts of natural images usually encode detailed information, while low-frequency parts often encode global structures. Therefore, mapping the output features of convolutional layers into high-frequency and low-frequency parts may enhance feature learning with multi-granularity. In this paper, we decompose the output features from convolutional layers into high-frequency and …