| The rapid development of SAR imaging technology has accumulated massive image data.How to extract information efficiently from massive data has gradually become a research hotspot.SAR object detection aims to automatically extract features from images to locate and classify targets.It can provide key technical support for engineering applications in the fields of disaster monitoring,marine resources monitoring,military target reconnaissance,agricultural forestry resources planning and exploration.In recent years,the deep learning technology represented by convolutional neural network has been widely used.SAR object detection based on convolutional neural network has become the mainstream technical.The convolutional neural network constructs the depth model by stacking multiple types of convolution layers,pooling layers and activation functions,and uses iterative training to automatically learn the general characteristics of the target from massive data,so as to realize the intelligent interpretation of the data.Therefore,facing the task of SAR image target detection,this paper carries out the research on SAR image target detection technology based on deep representation learning from the two aspects of SAR rectangle object detection and SAR oriented object detection.The specific work as follows:(1)Anchor-free SAR object detection based on multi-scale feature enhancement learning: due to the unique characteristics of SAR image,such as small sample,strong scattering,sparsity,multi-scale,complex background and unclear target edge contour,the existing SAR object detection methods directly migrate the mature target detection algorithm in the field of optical image to the target detection task of SAR image,which has limited performance and is difficult to balance the accuracy and speed.Therefore,we proposed an anchor-free algorithm based on multi-scale feature enhancement learning,which is called AFSar for short.Firstly,the latest anchor-free frame detection model Yolox is introduced as the basic framework;Secondly,in order to reduce the computational complexity and improve the ability of multi-scale feature extraction,we design a new lightweight backbone,namely Mobile Netv2S;Furthermore,we propose an attention enhancement PAN module,called CSEMPAN,which integrates channel and spatial attention to highlight the unique strong scattering characteristics of SAR targets;Finally,according to the characteristics of multi-scale and strong sparsity of SAR targets,we propose a new target detection head,called ESPHead.ESPHead uses convolution modules with different dilated rates to improve the receptive field,so as to enhance the ability of the model to extract the significant information of multi-scale SAR targets and further improve the detection performance.The comparative experimental results based on large-scale SAR image target detection data set SSDD show that compared with the traditional SAR target detection algorithm and the latest single-stage and two-stage target detection algorithm,our AFSar algorithm has higher SAR target detection accuracy,and the map reaches0.977;At the same time,our algorithm has lower computational complexity,and flops is only 9.86 G.(2)SAR oriented object detection based on deep feature alignment learning: the target orientation of SAR image in large scene is different,and the traditional object detector has some difficulties,such as rotation invariance,lack of feature information,time-consuming generation of directional candidate frame and so on,resulting in limited SAR oriented object detection efficiency.Therefore,this paper proposes a SAR oriented object detection algorithm based on deep feature alignment learning,which is called FADet.Firstly,we introduce the latest detection framework Oriented RCNN as the infrastructure;Furthermore,according to the characteristics of rotating targets,we introduce a feature alignment backbone Re Resnet to extract the rotation invariant features of SAR targets;Finally,we replace FPN with Re FPN to adapt to the semantic features by the backbone network;further,we propose a new RPN,namely AO RPN,we introduce an ARF module in the Head area of AO RPN to further improve the target rotation invariance;Finally,Ri Ro I Align is introduced,Ri Ro I Align consists of two parts: spatial alignment and orientation alignment.For RRo Is,spatial alignment refers to mapping from rotated features to rotation-invariant features in spatial dimensions,and orientation alignment refers Ri Ro I Align that also can maintain orientation alignment in the orientation dimension,and Ri Ro I Align produces completely rotation-invariant features from both spatial alignment and orientation alignment.The comparative experimental results based on the large-scale SAR image rotating target detection data set SSDD+ show that: compared with the latest single-stage and two-stage oriented object detection algorithms,the FADet algorithm in this paper has higher SAR target detection accuracy.The Recall,F1 and m AP reached 0.899,0.952,0.925 and 0.902,respectively.The research content of this paper can provide key technical support and theoretical basis for engineering applications in military and civilian fields such as aircraft /Satellite / missile borne SAR image intelligent interpretation. |
