Signal, Image and Video Processing Обработка сигналов
Prototype-Anchored prompting for few-shot remote sensing image classification
Аннотация
In few-shot learning for remote sensing image classification, the dual constraints of sample scarcity and intra-class variability lead to unstable class prototypes and high susceptibility of prompts to overfitting. To address this issue, this paper proposes a few-shot remote sensing classification framework called Prototype-Anchored Prompting(PAP): firstly, the Prototype-Anchored Gaussian Ensembling (PAGE) performs Gaussian-weighted aggregation on multi-round prompts during training, with weights biased toward the middle training phase, which balances adaptability and generalization and forms a cross-epoch global class center, alleviating the observed prototype drift phenomenon. Secondly, Textual Diversity Regularization (TDR) constructs a textual prototype pool using multiple templates and jointly constrains prompt learning with image-side prototypes, which effectively mitigates the modal imbalance of “diverse images versus sparse text” and enriches the semantic expression of class centers. Finally, Prototype Distillation Consistency Loss(PDC) imposes consistency constraints on both feature representations and prediction results at the textual and image ends, which aligns the prompt-enhanced model with the frozen CLIP at two levels and ensures that the class semantic prototypes in the training set maintain good discriminability and generalization. Experimental results demonstrate that PAP achieves more stable prototype estimation and stronger cross-epoch generalization, and significantly outperforms existing methods on multiple remote sensing benchmarks.