Research On Target Detection And Discrimination For SAR Images Based On Convolutional Neural Network

Synthetic Aperture Radar(SAR)is not affected by weather,light and other conditions and can work in all day.Thus,it has been widely applied in military and civilian fields.And SAR automatic target recognition(ATR)is a critical way for SAR image interpretation.The SAR ATR consists of three stages: target detection,target discrimination,and target classification.As the foundation of the SAR ATR,the performance of the SAR target detection and discrimination stages will affect the target classification stage.Therefore,research on the SAR target detection and discrimination is of great significance.Recently,deep learning methods have developed rapidly in the field of optical image processing and achieved better results.To solve the problems of target detection and discrimination for SAR images in complex scenes,this dissertation makes an in-depth research on the SAR target detection and discrimination by combining with the deep learning methods.The main contents of the dissertation are summarized as follows:1.We firstly introduce the research backgroud and signification of the dissertation,the development and key issues of the SAR target detection and target discrimination,and the main contents of subsequent chapters of the dissertaion.2.We study the traditional SAR target detection method and SAR target discrimination method.For the traditional SAR target detection method,firstly,the theoretical basis and implementation process of the constant false alarm rate(CFAR)detection algorithm is introduced.And then,the two-parameter CFAR and the chip extraction method based on the superpixel segmentation and the two-parameter CFAR are introduced in detail.For the traditional SAR target discrimination method,the image preprocessing,the traditional discriminative feature extraction,feature selection,and discriminator are described.The performance of the two-parameter CFAR and the traditional SAR discrimination features are analyzed by the experimental results on the measured SAR data.3.To solve the problem that the performance of the traditional SAR discrimination features decrease when the segmentation result for target and background is not ideal or the target and clutter do not have obvious difference in texture,size,and contrast,and so on,we propose the feature-fused SAR target discrimination method using multiple convolutional neural networks.In the proposed method,we utilize the stronger feature representation power of the convolutional neural network(CNN)and the intensity and edge information of SAR images.Specifically,the designed two-channel CNN architecture is applied to extract the deep features from the input SAR intensity images and the gradient amplitude images.Then,the deep features are fused into a new discriminative feature by using the proposed fusion approach.The proposed fusion approach can preserve the spatial relationship of the different features and effectively improve the discrimination performance of the fused feature.Finally,the new feature is fed into the classification network.The experimental results on the measured SAR data demonstrate the effectiveness of the proposed method.And the proposed method performs better than the traditional discrimination methods,especially in the complex SAR scene.4.For the case of the limited labeled training data,we propose a semi-supervised SAR target discrimination method.CNN requires a large number of labeled training data to ensure its better performance.And the collection of the labeled SAR data is expensive and time demanding.It may yield overfitting when directly training CNN with the limited labeled data.The proposed method contains two components,i.e.,the classification network and the reconstruction network,which can simultaneously use a large number of unlabeled data and a limited number of labeled data to extract generalized feature.Because the classification network and the reconstruction network share part of the whole structure,the feature about the essential knowledge of the data can be extracted by reconstructing the unlabeled data,which can provide the useful information to the feature extraction process of the classification network.In addition,the scenes of SAR images are complex.The type,location and orientation of the target are varying.The clutter type is diversified.The above problems make the distributions of the training data and test data be different.The performance of the network is limited due to the different distributions of the training and test data.To solve this problem,we propose a feature constraint term based on Kullback-Leibler(KL)divergence,which enforces the distributions of the learned feature of the training and test data to be close to each other in the feature space.The experimental results on the measured SAR data demonstrate the effectiveness of the proposed method with the limited labeled training data.5.Aiming at the problem of imbalance between positive and negative samples in SAR target detection based on deep learning,we propose a CFAR-guided EfficientDet with improved loss function for SAR target detection.Firstly,the EfficientDet,which achieved much better efficiency in the field of optical image processing,is applied to realize target detection for SAR images.Since the labeled SAR data is limited and the SAR scene contains only a small number of target areas,the problem of imbalance between positive and negative samples is more serious in the network training process.To solve this problem,we improve the EfficientDet from two aspects,i.e.,the loss function and the process of generating candidate boxes.On the one hand,the Average-Precision loss(AP-Loss)that is less effected by the ratio of the number of positive and negative samples is introduced into the loss function of the classification branch,thereby improving the detection performance of the network.On the other hand,because the SAR image targets have the characteristics of strong scattering,the clutter areas that are easy to distinguish in the original SAR image can be removed by the two-parameter CFAR.Thus,we effectively integrate the binary image obtained from the two-parameter CFAR into the network to guide the generation process of the candidate boxes.Through this way,the number of negative samples that are generated by the network and easy to distinguish is greatly reduced,thereby further suppressing the issue of imbalance between positive and negative samples.The experimental results using the measured SAR data demonstrate that the proposed method has better detection performance and can effectively alleviate the problem of imbalance between positive and negative samples.