Multimodal Feature Calibration Network for Few-Shot Remote Sensing Image Scene Classification

Existing few-shot remote sensing scene classification methods that utilize semantic textual information primarily classify by fusing image information of the support set with its semantic text. Although this approach has led to some performance improvements, it lacks rationality due to the information discrepancy between the support and query sets. To address this issue, a feature calibration mechanism utilizing multimodal information is proposed. For the support set, features are calibrated in two ways to make the image features closer to the real prototypes: using semantic textual information and local view information. For the query set, although it lacks semantic information, the feature calibration relies on local view information, effectively mitigating the impact of information imbalance. Finally, guided by the calibrated query set, an attention mechanism is employed to fuse the two calibrated support set features into a new prototype feature for classification. Experimental results demonstrate that the proposed method significantly improves classification accuracy compared to existing methods.