SAR image target recognition algorithm based on CN-GAN and CNN

Abstract

The invention discloses an SAR (Synthetic Aperture Radar) image target recognition algorithm based on a CN-GAN (Convolutional Neural Network) and a CNN (Convolutional Neural Network), and the method comprises the following steps: S1, the CN-GAN is combined with a least-squeries GAN and Pix2Pix method, and a regression function constraint term is added to a loss function of a generator, so that theimproved CN-GAN can solve data generation problem; and S2, the CNN directly takes the SAR image generated by the CNGAN as network input, a shallow network design is adopted in structure, and a Dropout layer is added, so that the generalization ability of the model is effectively improved. The recognition rate of the SAR image data set generated by the multiplicative noise-containing data set through the CN-GAN is superior to the recognition rate of the multiplicative noise-containing data set and the recognition rate of the data set generated by other GANs.

Description

technical field [0001] The present invention relates to the technical field of target recognition for synthetic aperture radar (SAR) images, in particular to a Constrained Naive Generative Adversarial Networks (CN-GAN) and Convolutional Neural Network (Convolutional Neural Network), CNN) SAR image target recognition algorithm. Background technique [0002] Affected by complex environments and special imaging principles, Synthetic Aperture Radar (SAR) images still have the problem of sample scarcity and large coherent speckle noise, which easily affects the effect of target recognition. [0003] In recent years, the technology of SAR has been developing faster and faster, the image quality is getting better and better, and the image resolution is getting higher and higher, but the development of Automatic Target Recognition (ATR) based on SAR images is relatively slow. The difficulties of SARATR are mainly reflected in three aspects: (1) The influence of complex environment,...

AI Technical Summary

Problems solved by technology

This method solves the problems of the conventional naive GAN with large output degrees of freedom, the high training complexity of the constrained GAN model, and the sharp drop in performance in low signal-to-noise ratio environments.

Images