Efficient SAR Automatic Target Recognition Based on Self-distillation

Many current SAR-ATR methods utilizing deep learning mainly emphasize improving model generalization. Consequently, this approach results in a substantial expansion of the model’s parameter set. This study introduces a new neural network architecture that combines low parameter count with high generalization, incorporating a self-distillation technique during the training process. Initially, we employ a substantial quantity of DWConv to diminish the count of trainable parameters and merge features with distinct receptive fields to enhance model efficacy. During the training phase, the model backbone’s generalization is enhanced through the implementation of a self-distillation technique, which is not incorporated in the inference process. Experimental evaluations conducted under the OFA protocol reveal that the proposed method achieves outstanding performance and generalization.