Image Classification with Frequency Channel Attention under the Few-Shot Condition

Few-shot learning aims to use very few samples for classification. In the existing few-shot learning methods, most of them only use the spatial information of images. In few-shot learning, because of the lack of samples, it is very important to improve the expression of image features. Recent studies have found that the feature representation in frequency domain contains a wealth of information can be used for tasks image classification. Therefore, this paper proposes to use the frequency domain information of images to enhance feature representation to improve the utilization efficiency of a small number of samples. We propose a network called Meta-Baseline-Fca. It takes a simple small-sample learning network Meta-baseline as the baseline, adds a frequency channel attention (Fca) module in the network, and uses the spatial and frequency domain features of the input image to enhance the input image features expression. We conduct comparative experiments in the miniImageNet dataset and tiereimageNet dataset. The results show that under the condition of small samples, the Fca module can significantly improve the classification accuracy. Compared with the Meta-Baseline network, the classification accuracy is improved by 0.92% in the miniImageNet dataset and 0.84% in the tiereimageNet dataset.