Spatial-frequency attention and fusion network for few-shot fine-grained image classification

The objective of the few-shot fine-grained image classification task is to categorize image data using only a limited number of labeled samples. These photographs typically pertain to various subclasses within the same species and exhibit minor visual distinctions. When classifying these images, accurately capturing distinctive features that are useful for classification is of utmost importance. Therefore, this paper designs a Spatial-frequency attention and fusion network (SFAFNet) for few-shot fine-grained image classification. Specifically, the method not only extracts features in the spatial domain but also converts the image to the frequency domain to filter out noise interference. A network employing non-shared parameters is utilized to derive its frequency domain feature representation, thereby obtaining diverse and significant feature representations across various spaces. Secondly, multi-level information of each branch is extracted to obtain more accurate significant features. Finally, using the double unique quantity module, measurement and weighted fusion are performed respectively in different spaces to obtain more accurate results. Experiments show that it has been verified on multiple fine-grained benchmark datasets. The method introduced in this paper outperforms other advanced algorithms.