Cross-Domain Few-Shot Fine-Grained Classification Based on Local-Global Semantic Consistency and Earth Mover’s Distance

In recent years, few-shot classification algorithms based on metric learning have gained significant attention in the field. However, in the context of cross-domain few-shot classification tasks, their performance still requires further improvement. To address this limitation, this paper proposes a cross-domain few-shot fine-grained classification model based on local-global semantic consistency. In order to tackle cross-domain few-shot classification challenges, we introduce a cross-loss computation method in this model. This method leverages the differences and similarities between global views and local views of each image, enabling the model to learn the local-global semantic consistency of the images. Through this approach, the model can capture shared features across different domains, thereby reducing intra-feature semantic differences and enhancing the consistency of class relationships in sample predictions. Building on the comprehensive utilization of local and global features in images, our model excels at focusing on local details closely related to specific categories. This further strengthens intra-class associations among similar images, leading to more effective discrimination of fine-grained categories. In summary, our proposed cross-domain few-shot fine-grained classification model, based on local-global semantic consistency, not only aims to address the challenges of metric learning in cross-domain few-shot classification but also provides a promising approach for fine-grained classification tasks.