NR-MAE: noise reduction masked autoencoder for boosting SSL on SAR target recognition

Self-supervised learning models can effectively adapt to the prevailing big data trend, leading to their extensive applications in various domains. Nevertheless, image degradation caused by speckle noise leads to a performance decline of self-supervised learning models in Synthetic Aperture Radar Automatic Target Recognition (SAR ATR). In this article, we introduce a Noise Reduction Masked Autoencoder (NR-MAE) method that integrates the Masked Image Modeling(MIM) with filtering algorithms. The NR-MAE reduces speckle noise effects through filtering algorithms, enabling the model to learn target features more accurately and efficiently. Experiment indicates that NR-MAE outperforms the MAE method in SAR image recognition accuracy, showing an improvement of 8% to 10% in performance under the 1/32scale small sample training set. Furthermore, an ablation study on the filtering algorithm's hyperparameters affirms the model's robustness to changes in these parameters.