SAR Image Target Classification And Detection Based On Lightweight Deep Learning

Synthetic Aperture Radar(SAR)is one of the main methods of SAR image acquisition.As an active microwave device,it mainly uses the back-scattering of objects for imaging,and the formed SAR image target has the characteristics of unclear outline information,complex background and strong scattering.Based on this,the SAR image target classification and detection technology is derived,and the research on SAR image target classification and detection in the direction of remote sensing image interpretation has great theoretical value and engineering application significance.Since deep learning has shown excellent performance in the direction of artificial intelligence interpretation,the reason is its powerful feature representation ability.In this context,this dissertation focuses on the two core tasks of SAR image target classification and SAR ship detection,conducts research on SAR image target classification and detection technology based on lightweight deep learning.The main research contents and innovative work of this dissertation are as follows:(1)In view of the difficulty of balancing the classification accuracy and speed of the existing SAR image target classification methods,and the unique characteristics of SAR image targets are not considered in depth,this dissertation proposes a SAR image target classification algorithm based on lightweight deep learning,referred to as Sar-Efficient Net.First of all,this dissertation introduces the latest Efficient Net-B0 in the field of computer vision as the basic network model,which comprehensively weighs the depth,width and image resolution of the model,and is more suitable for SAR image target classification;Secondly,in order to reduce the computational complexity of the model,this dissertation designs a lightweight backbone,called LWbackbone,it only retains the first SE module in the mobile reverse bottleneck MBConv,and simplifies the stacking part of MBConv;Further,in order to improve the classification accuracy of the model,this dissertation uses a depthwise separable convolution module is incorporated into the MBConv structure.The comparative experimental results based on the typical SAR image target classification data set MSTAR show that compared with the existing latest SAR image target classification algorithms,the Sar-Efficient Net algorithm proposed in this dissertation has higher classification accuracy,and the accuracy rate reaches 99.5%;The algorithm in this dissertation has lower computational complexity and parameter amount,which are reduced to 277.56 M and 4.01 M,respectively,on the premise of ensuring the improvement of accuracy.(2)Most of the existing deep learning SAR image ship detection algorithms based on anchor frame rely on expert experience to set a series of hyper-parameters,and the generalization performance is weak;It is difficult to characterize the unique characteristics of SAR image ship targets,which greatly reduces the accuracy and speed of detection.To this end,this dissertation proposes an anchor-free SAR ship target detection algorithm based on lightweight position-enhanced representation,called LPEDet.First,in order to solve the problem of unclear outline and multi-scale SAR target,this dissertation introduces YOLOX as the benchmark framework,and designs a lightweight multi-scale backbone network,called NLCNet,to balance detection accuracy and speed;Due to the strong scattering characteristics,this dissertation designs a position-enhanced attention strategy,which suppresses background clutter by adding position information to the channel attention that highlights the target information,so as to identify and locate the target more accurately.The comparative experimental results based on the large-scale SAR ship target detection data sets SSDD and HRSID show that the method in this dissertation has the best SAR target detection results,with m AP reaching 97.4%/89.7% respectively;The speed is significantly improved,and FLOPs and the average inference time for single image drops to18.38G/28.72 G and 7.01ms/7.11 ms,respectively.The research content of this dissertation can provide key technical support for the field of intelligent interpretation of SAR images,and provide important engineering solutions for the development of multi-platform combat applications such as air-borne,space-borne,and missile-borne.