Study On Target Detection And Target Recognition For SAR Images Based On Multi-Model Joint Learning

As an active microwave sensor,synthetic aperture radar(SAR)can form high resolution images with relative invariance to weather and lighting conditions.Compared with infrared sensors,optical remote sensing and other passive sensors,SAR cannot be affected by some environmental factors such as light conditions and climate change,and has the penetration ability to some extent.Therefore,it has been widely used in military and civilian fields.As the important parts of SAR image interpretation,target detection and target recognition have received extensive attention from scholars at home and abroad.However,due to the poor interpretability of SAR images,it is difficult to label and extract separable features.A single simple model often fails to obtain satisfactory detection and recognition results.This thesis studies the SAR target detection and recognition method based on multi-model joint learning,and obtains the improvement of detection and recognition performance through joint learning of multiple single models,and proposes two SAR target detection methods and two SAR target recognition method respectively for the above problems.The main contents can be summarized as the following three aspects:1.Aiming at the problem that it is difficult to obtain labeled training samples in the SAR target detection task,two unsupervised SAR target detection methods are proposed in the third chapter of the thesis based on the single-source domain and the multi-source domain.First,a Domain Adaptive Single Shot Multibox Detector(DASSD)model is established based on single-source domain and multi-domain joint learning.In DASSD,a Generative Adversarial Network(GAN)constraint is used to jointly learn the source domain optical image target detection model and the target domain SAR target detection model,and the source domain target detection model is composed of a large number of The knowledge learned from the labeled data is effectively transferred to the target detection model in the target domain to realize unsupervised SAR target detection.For the problem of domain migration,one is to use the gradient reversal layer to share the feature extraction layer of the source domain detection model and the target domain detection model.The second is to design an iterative pseudo-label strategy to fine-tune the model after the domain adaptation learning is completed.Further,in view of the problem that DASSD cannot use all source domain information at the same time when there are multiple source domain data sets with different distributions,a Multi-source Domain Adaptive SSD(MDASSD)model is established.First,each source domain data set is trained to adapt the SSD target detection sub-model to the target domain SAR data knowledge transfer domain.Secondly,in the stage of fine-tuning the model using pseudo-labels,in order to alleviate the problem of inaccurate pseudo-label marking,the GAN distance is designed as a criterion to select samples in the source domain that are closer to the target domain to also participate in the model fine-tuning.Finally,an SSD-based decision-level fusion method is established,and the results of each domain adaptation knowledge transfer sub-model from the source domain to the target domain are fused at the decision-making level.Based on the measured data,it is verified that the performance of DASSD is much higher than the traditional unsupervised SAR target detection method.At the same time,compared with the single-source domain adaptation model,the MDASSD model can effectively use multiple source domain information and obtain better SAR target detection results.2.It is difficult to extract separable features from SAR data.In addition,the traditional factor analysis(FA)model ignores the two-dimensional spatial correlation of pixels in SAR data.To solve the problem,a max-margin multi-scale convolutional factor analysis model(MMCFA)is proposed in the fourth chapter of the thesis.Different from FA model,MMCFA model introduces two-dimensional convolution operator,which can directly learn the features of two-dimensional SAR image pixels in two-dimensional space.And by setting multi-scale convolution kernel,MMCFA model can achieve multi-level rich feature extraction.Furthermore,the multi-scale convolutional FA model is combined with a latent variable representation of support vector machine(LVSVM)to improve the separability of hidden features.The experiments based on MSTAR dataset show that MMCFA model has better recognition performance than traditional recognition methods.3.The traditional denoising methods in noise robust SAR automatic target recognition research are independent of the recognition model,which limits the robust recognition performance.Combined with deep learning method,we present a robust SAR automatic target recognition method via adversarial learning,discriminative dual generative adversarial network(DDGAN),which could integrate data denoising,feature extraction and classification into a unified framework for joint learning.Different from the common recognition methods of directly inputting the SAR data into the classifiers,we add a dual generative adversarial network(DGAN)model between the SAR data and the classifier for data translation from a noise-polluted style to a relatively clean style to reduce the noise from SAR data.In order to ensure the target information in the SAR data can be retained during the data style translation,reconstruction constraint and label constraint are also used in the dual GAN model.Then the more reliable transferred SAR data is fed into the classifier.The parameters of the dual GAN and classifier are learned through joint optimization in our method.Thus the data separability is guaranteed in the process of denoising and feature extraction,which greatly improves the recognition performance of the method.Experimental results on MSTAR dataset show that the proposed DDGAN model has better recognition performance than traditional noise robust recognition methods in the case of low signal-to-noise ratio.