The Deep Network Design For SAR Target Detection And Recognition Combining With Optical Knowledge And Electromagnetic Scattering Characteristics

Synthetic Aperture Radar(SAR)is an active microwave sensor,which uses pulse compression technology and synthetic aperture technology to realize two-dimensional high-resolution imaging.SAR is not only capable of all-weather and all-day observation of the earth,but also has a certain penetration ability,which has more advantages than optical,infrared and other sensors.As important steps in SAR image interpretation,SAR target detection and recognition have received widespread attention and in-depth research.With the development of deep networks,due to their powerful feature learning capabilities and significant performance advantages,they have gradually become the mainstream method for research in the field of SAR target detection and recognition.However,existing deep neural network-based SAR target detection and recognition methods rely on a large number of labeled training samples.Due to the high difficulty of interpreting SAR images,it is extremely challenging to accurately label SAR images.Meanwhile,compared with optical images,there are fewer SAR images available for model training.In addition,most of the existing deep learning models for SAR target detection and recognition directly use or simply improve the optical image perception models,and do not make full use of the unique electromagnetic scattering characteristics of SAR targets.All these problems have greatly affected the performance improvement and popularization of SAR target detection and recognition networks.Therefore,with the help of the theory and framework of deep learning,this dissertation analyzes and researches the design of deep networks for SAR target detection and recognition combining with the optical knowledge and electromagnetic scattering characteristics.The main contents of this dissertation are summarized as follows:1.Since it is difficult to acquire the labels of SAR images while it is easier to acquire and label optical images,utilizing the useful information in optical images to assist the target detection of SAR images for similar scenarios is a feasible idea.Therefore,Chapter 3proposes an unsupervised domain adaptation deep network for SAR target detection combining with the optical knowledge.The proposed network aims at improving the target detection performance of unlabeled SAR images with the help of the useful information in optical images by designing a suitable network structure.There are two stages in this network.In the first stage,the idea of feature decomposition is first used to better extract the domain-shared features between optical images and SAR images,and then the domain-shared features extracted from optical images and their labels are used to train the domain-shared detector.In the second stage,the uncertainty-guided self-learning method is used to further extract the discriminative features of SAR images based on the trained domain-shared detector.The extracted discriminative features of SAR images are used to train the dedicated detector for SAR images,thereby avoiding the loss of some unique information of SAR images that may be valuable for detection.Experiments based on measured SAR images demonstrate the network designed in this chapter can significantly improve the target detection performance of unlabeled SAR images with the assistance of fully labeled optical images.2.Most of the existing SAR target recognition methods based on Convolutional Neural Network(CNN)directly use or simply improve the optical image perception models,and do not make full use of the unique electromagnetic scattering characteristics of SAR targets,which leads to the limited descriptive capability of the features.To solve the problem,Chapter 4 proposes a deep network for SAR target recognition combining with the Attributed Scattering Center(ASC)schematic maps.The proposed network aims to enhance the feature description ability by designing a suitable network structure combining with the electromagnetic scattering characteristics of SAR targets,so as to improve the performance of SAR target recognition.The network contains two different inputs: one is the real-valued SAR image,the other is the ASC schematic map that reflects the physical structure information of the SAR target,both of which are obtained by processing the original complex-valued SAR data.In addition,the network designs two parallel feature extraction branches with different network structures based on the characteristics of the two types of data.They are responsible for extracting more discriminative image features and physically meaningful features from the input real-valued SAR images and ASC schematic maps,respectively.In order to obtain features with stronger descriptive capabilities to improve target recognition performance,the network uses a feature fusion module to fuse the two obtained features,and uses the fused features to predict the category of the input image.Experiments based on measured SAR images demonstrate the designed network can considerably improve the performance of SAR target recognition and alleviate the problem that training samples of SAR images are limited to a certain extent.3.The ASC schematic maps used in Work 2 have some problems such as insufficient separability and not fully corresponding to the targets in real-valued SAR images.To solve these problems,Chapter 5 proposes a deep network for SAR target recognition based on ASC component analysis.The proposed network aims at utilizing the ASC information of the SAR target more effectively to enhance the feature description ability by designing a suitable network structure,thereby further improving the performance of SAR target recognition.Considering that the component information of the target is robust to local variations,and the ASC can reflect the real physical structure of the SAR target,this chapter first divides SAR targets according to the scatterer types of ASCs to obtain component division results that are not only more robust but also can accurately characterize the electromagnetic scattering characteristics of the target.Further considering that the overall information and shadow information of the target contained in the global information provided by the whole image are also important,the network combines the global information with the component information to obtain more robust and descriptive features.In addition,the network also utilizes the fusion of multi-scale feature maps to enhance the descriptive capability of the global feature maps,which further enhances the descriptive capability of the final fused features.Experiments based on measured SAR images demonstrate the network designed in this chapter can not only further improve the SAR target recognition performance,but also improve the recognition accuracy of variants.At the same time,it can alleviate the problem that training samples of SAR images are limited to a certain extent.4.Aiming at the problems that existing CNN-based SAR target recognition methods usually rely on a large number of training samples,as well as the extra parameters and network complexity brought by the operations such as extracting ASCs for each SAR image and performing multiple feature fusions in Work 3,Chapter 6 proposes a deep network for SAR target recognition based on ASC convolutional kernel modulation.The proposed network aims at designing a simpler network structure combining with the electromagnetic scattering characteristics of SAR targets to extract features that are more in line with the characteristics of SAR images,so as to reduce the number of network parameters while ensuring its feature extraction capability.Considering that SAR images have electromagnetic scattering characteristics,in order to extract linear structures and edge features that are more consistent with SAR target characteristics,a small number of CNN convolutional kernels are modulated using predefined ASCs with different orientations and lengths to generate more convolution kernels,which can reduce the parameters of network while ensure its feature extraction capability.The designed network uses ASC modulated convolutional kernels at shallow layers to extract linear structure and edge features that are more in line with the characteristics of SAR images,while using CNN convolutional kernels at deeper layers to extract semantic features of SAR images.Due to the simultaneous use of ASC modulated convolutional kernels and CNN convolutional kernels,the proposed network is able to give attention to both the electromagnetic scattering characteristics of SAR targets and the feature extraction advantages of CNN.Experiments based on measured SAR images demonstrate the network designed in this chapter can ensure excellent SAR target recognition performance while reducing the demand for training samples.