Self-Supervision and Self-Distillation with Multilayer Feature Contrast for Supervision Collapse in Few-Shot Remote Sensing Scene Classification

Although the means of catching remote sensing images are becoming more effective and abundant, samples that can be collected in some specific environments quite scarce. When there limited labeled samples, methods for analyzing scene classification perform drastically worse. Methods classify few-shot image scenes often based on meta-learning algorithms handling sparse data. However, this research shows they will affected by supervision collapse where features help with out-of-distribution classes discarded, which is harmful generation unseen new tasks. In work, we wish to remind readers existence propose a method named SSMR multi-layer feature contrast overcome collapse. First all, makes use label information contained finite number guides self-supervised learning train embedding network generated multilayer contrast. This prevent from losing intra-class variation. Intra-class variation always useful classifying What more, merged self-distillation, modified self-distillation used encourage extract sufficiently general transfer better domains. We demonstrate most existing suffer overcomes well experiments dataset specially designed examining problem, 2.4–17.2% increase compared available methods. Furthermore, performed series ablation how necessary each structure proposed show different choices training impact final performance.