Attribute- and attention-guided few-shot classification

The field of image classification faces significant challenges due to the scarcity of target samples, leading to model overfitting and difficult training. To address these issues, few-shot learning has emerged as a promising approach. However, current methods do not fully utilize the correlations among samples and external semantic information, resulting in poor recognition accuracy. To overcome these limitations, we propose a new few-shot classification method that incorporates both attributes and attention guided approach. The method leverages the attention mechanism to extract discriminative features from the images. By exploring regional correlations among samples, it assists in generating visual representations by utilizing predicted attribute features. As a result, accurate prototypes are generated. Extensive experiments were conducted on two attribute-labeled datasets, namely Caltech-UCSD Birds-200–2011(CUB) and SUN Attribute Database (SUN) Attribute Dataset. With the Resnet12 backbone, the method achieves remarkable accuracies of 79.95% and 89.34% for 1-shot and 5-shot, respectively, on the CUB dataset. Similarly, with the Conv4 backbone, the method achieves notable accuracies of 67.21% and 80.87% for 1-shot and 5-shot, respectively, on the SUN Attribute dataset. The achieved accuracies highlight the robustness and generalizability of our method, and show the capability of our method to accurately classify samples with limited training data, which is a significant advantage in real-world scenarios where labeled data are often scarce.