Self-attention multiscale feature fusion network for small sample SAR image recognition

Synthetic Aperture Radar (SAR) image tags were difficult to obtain in large quantities, and there were large numbers of small sample SAR datasets. SAR image was full of scattered noise, and it was difficult to extract effective features by directly applying Convolutional Neural Network (CNN) on small sample SAR datasets. To addressed the above problems, this paper proposed a self-attention multiscale feature fusion network for small sample SAR image recognition. Firstly, the self-attention mechanism was combined with the ghost module to construct the self-attention ghost module, and the module was used to replace the classic convolution operation to extract SAR image features. Secondly, a channel shuffle unit was added to the network to construct multiscale information fusion branchs. Finally, knowledge distillation was introduced to compress the designed network and further control the number of network parameters. The experimental results show that the method in this paper has excellent recognition performance on MSTAR datasets collected under different operating conditions and also exhibits good robustness on small sample SAR datasets constructed.