| Synthetic Aperture Radar(SAR)is a high-resolution sensor that is not easily affected by natural factors,and can achieve earth observation in all-weather and all-day.It has been widely used in military,agriculture,forestry,ocean,disaster monitoring and other fields.With the rapid development of SAR imaging technology,the quantity of obtained SAR image is increased,and the quality of available image is improved significantly,which brings greater challenges to SAR image interpretation.SAR automatic target recognition(ATR)is a critical technology for SAR image interpretation,and it is mainly composed of three stages:target detection,target discrimination and target classification.The SAR target detection and target discrimination stages are the foundation of the entire SAR ATR system,and their performance will directly affect the results of subsequent target classification.How to quickly and accurately detect targets from SAR images is still a current research hotspot.In recent years,deep learning has developed quickly in the field of optical images with powerful feature extraction capabilities.To solve the problems of SAR image target detection,this paper combines the traditional constant false alarm rate with deep networks.The main research contents of this dissertation are summarized as follows:1.A step-by-step target detection method combining constant false alarm rate(CFAR)detection and target discrimination is studied.Firstly,the theoretical basis of the CFAR detection and the detection method using global threshold is introduced,then the SAR target discrimination method based on feature selection is introduced in detail.On the measured Mini SAR image data,the suspected target area is obtained from the SAR image through the detection algorithm,and the advantages and disadvantages of the two detection algorithms are compared;Then,the traditional SAR target discrimination method based multiple feature selection and the SAR target discrimination network based on convolutional neural network(CNN)are used to identify the obtained suspected target chips,and the final SAR target detection result is obtained.Compared with the traditional SAR target discrimination method,the SAR discrimination method based on CNN network has stronger ability to discriminate SAR image chips in complex scenes.2.The end-to-end object detection algorithm based on deep network is studied.Firstly,two classic target detection algorithms based on deep network are introduced in detail: Faster RCNN and Single Shot Multibox Detector(SSD).And SSD is applied to the SAR image target detection task in complex scenes to achieve the end-to-end integrated target detection;Then,the SSD detection network is improved for the specific ground vehicle target detection task in SAR image.Experiments on the measured Mini SAR image data show that the SAR target detection algorithm based on the deep network can obtain better detection results than the traditional step-by-step target detection method.3.Aiming to solve the problem of SSD algorithm when applying it to the SAR target detection task,this paper proposes a SAR target detection algorithm combining CFAR detection algorithm and SSD.Firstly,in view of the insufficient feature extraction capability of the SSD,a multi-scale feature fusion liking path aggregation network,i.e.PANet network,and a convolutional block attention module(CBAM)are introduced into the SSD feature extraction network;Then,in order to solve the problem of insufficient utilization of SAR image characteristics by SSD,based on two-parameter CFAR detection,the characteristic of strong scattering of SAR target is fully utilized through the dual-stream sub-network structure and the interactive channel-spatial attention fusion(ICSAF)module.Then,aiming at the problem of imbalance between positive and negative samples in SAR target detection based on deep learning,this paper proposes a new classification loss function based on CFAR detection and Focal Loss function.Finally,for the post-processing problem of large scene SAR images,a new non-maximum suppression algorithm based on area-ratio(ARNMS)is proposed.The experimental results on the measured Mini SAR data demonstrate the effectiveness of the proposed method. |
