Target Recognition of Few-sample SAR Images Based on Style Migration and Feature Alignment

Convolutional neural network (CNN) has been widely used in the field of image recognition with its excellent performance, but in the problem of ground vehicle target recognition in SAR images, CNN models lack large measured data sets in the training stage, which is prone to overfitting phenomenon. At present, the electromagnetic simulation SAR image based on the target scattering characteristics is relatively easy to obtain, but there are significant domain differences between the simulation image and the measured image. Direct training of the simulation image occurs, which seriously affects the recognition performance. To solve this problem, this paper puts forward a use of electromagnetic simulation SAR image auxiliary training algorithm, through the combination of style migration and feature alignment, from the image of two levels step by step to implement the simulation image and measured image domain adaptation, make the feature extraction network can extract the common target classification features between the two. Experimental results on the SAMPLE dataset show that the proposed algorithm outperforms the classical domain adaptation algorithm and effectively improves the generalization ability of the recognition model.