A Small Sample Synthetic Aperture Radar Target Recognition Method Assisted By Other Source Information

Synthetic Aperture Radar(SAR)target recognition technology,through digital image processing and data analysis and interpretation,can extract target position,category,and other information from SAR images.It is one of the key technologies for SAR imaging systems to convert detection data into battlefield intelligence in military applications.Deep learning methods can automatically learn features from SAR data without the need for manual design and selection of features,reducing the cost and error of manual intervention.Currently,they have become the mainstream in SAR target recognition methods.However,SAR target recognition methods based on deep learning heavily rely on data.Due to the complex and changeable actual acquisition environment of radar,the acquisition of SAR images is difficult and the number of SAR images is small,which leads to the failure of the recognition algorithm to achieve the optimal performance.In recent years,many researchers have assisted in solving small sample problems through other sources of information,among which electromagnetic simulation technology is highly favored.It can artificially generate a large number of simulated SAR data,and achieve accurate control of target angle and other parameters,which can reduce the cost of data acquisition and increase the quantity.However,it was found that there are differences in electromagnetic simulation data compared to real samples,and the improvement in recognition performance is limited.In view of the above issues,this thesis focuses on image correction methods and sample selection strategies to effectively correct electromagnetic simulation data and reasonably select high-quality samples to expand the SAR image dataset.The main content and contributions include the following aspects:(1)In response to the significant difference between electromagnetic simulation images and real SAR images,which results in limited improvement in target recognition performance due to the direct use of simulation data,this thesis proposes an image correction method based on Texture Structure Cycle-consistent Adversarial Network(TS-Cycle GAN).Based on the analysis of SAR target characteristics,the loss function in Cycle GAN model is improved according to the texture and structure of the image and the structural similarity measurement index.This method combines the generative adversarial network model in unsupervised learning in the field of deep learning,and corrects the existing electromagnetic simulation image to a correction sample dataset with higher similarity to the real SAR image features,so as to improve the quality of supplementary data.(2)In response to the problem of uneven quality of corrected samples and high information overlap between some generated samples,this thesis proposes a sample selection strategy based on the Peri-center Clustering Selection(PCS)algorithm.Firstly,analyze the distribution characteristics of the corrected samples and the real samples in the feature space,select the corrected samples closer to the feature center of the real SAR samples to ensure the similarity between the screened samples and the real samples.At the same time,use the K-Means++ algorithm to avoid local clustering of the selected sample features to ensure the information diversity of the screened samples,thus achieving effective utilization of electromagnetic simulation data.