Self-Learning for Few-Shot Remote Sensing Image Captioning

Large-scale caption-labeled remote sensing image samples are expensive to acquire, and the training available in practical application scenarios generally limited. Therefore, caption generation tasks will inevitably fall into dilemma of few-shot, resulting poor qualities generated text descriptions. In this study, we propose a self-learning method named SFRC for few-shot captioning. Without relying on additional labeled external knowledge, improves performance by ameliorating way efficiency learning limited data. We first train an encoder semantic feature extraction using supplemental modified BYOL self-supervised small number unlabeled samples, where derived from samples. When model self-ensemble yields parameter-averaging teacher based integration intermediate morphologies over certain time horizon. The self-distillation uses self-ensemble-obtained generate pseudo labels guide student next achieve better performance. Additionally, when optimizing parameter back-propagation, design baseline incorporating self-critical reduce variance during gradient computation weaken effect overfitting. range experiments only evaluation metric scores exceed those recent methods. conduct percentage sampling test captioning with fewer also ablation key designs SFRC. results prove that these sparse sample indeed fruitful, each contributes method.