Research On Efficient Methods For High-resolution And Wide-swath SAR Imaging

Synthetic Aperture Radar(SAR)has the characteristics of all-day and all-weather,and is widely used in marine monitoring,disaster warning and other fields.High-Resolution and Wide-Swath(HRWS)SAR imaging can obtain high-resolution imaging while carrying out wide-format mapping,which is especially suitable for monitoring wide ocean areas.However,it takes a long time for traditional imaging methods to complete highresolution wide-format imaging,which cannot meet the practical application requirements of rapid detection of ships in sea areas,marine rescue,and monitoring of oil spills on the sea surface.In a wide and large scene area,the target area of interest is often relatively small.For example,compared with the vast ocean scene,the area of concern for ships and islands is small and sparsely distributed.For this reason,this thesis proposes a multi-level resolution imaging method:For large and wide target sparse scenes,the imaging results of the entire scene are quickly obtained at low resolution,and then the target area is quickly detected and accurately extracted,and then the target area is imaged at high resolution,thereby improving the imaging efficiency.The main contents and contributions of this thesis are as follows:1.A multi-level resolution SAR imaging method is proposed.The calculation amount of the traditional back projection algorithm is mainly determined by the number of spatial grids in the azimuth and distance directions and the number of azimuth sampling points.In order to reduce the calculation amount of the algorithm and improve the imaging efficiency,the number of azimuth sampling points can be directly reduced and the azimuth sampling point can be increased.The number of grids can be reduced by adjusting the grid spacing in the direction and distance direction,but it will reduce the imaging resolution.Combined with the target sparsity in the sea scene,this thesis proposes a multi-level resolution SAR imaging method.The basic idea is to first perform low-resolution fast imaging of the entire scene,then quickly detect the target area on the imaging results,and then perform high-resolution imaging of the target area.,thereby quickly obtaining high-resolution images of objects in large scenes.In order to quickly realize low-resolution imaging,the proposed method uses deoblique processing in the distance direction to obtain small bandwidth echoes,and uses sub-aperture imaging in the azimuth direction to reduce the amount of computation.By designing a special signal receiving method,the high-resolution echo data in the range direction and the low-resolution data after de-slope can be obtained at the same time,which avoids the multi-channel error correction and complex filter design problems caused by traditional downsampling processing.Conducive to saving radar hardware resources.The sub-aperture length in the azimuth direction refers to the sub-bandwidth resolution after de-slope in the range direction to obtain normal scale imaging results.Then use the constant false alarm detection method to detect and extract the low-resolution image,and then use the large-bandwidth full-aperture echo data to image the target area,and finally obtain a high-resolution target image.In addition,the complexity model of the multi-level resolution SAR imaging method is established,and the factors affecting the computational efficiency are analyzed.Finally,the echo data of point target,aircraft target and ship target under 1km×1km scene are simulated respectively,and the time and quality of direct full-scene imaging and multi-level resolution imaging are compared.The results show that the proposed method can maintain the target.At the same time of regional highresolution imaging,the imaging time is only 0.86%,0.906%,and 4.82%of the traditional BP algorithm,respectively.Moreover,the smaller the target area of the large scene,the more accurate the detected and extracted target area,and the more obvious the speed improvement.The simulation results verify the obvious advantages of the proposed method,and provide an effective way to quickly realize high-resolution imaging of large scenes.2.A low-resolution SAR image detection method based on knowledge distillation and probability distribution bounding box representation is proposed.The boundary of the target area in low-resolution SAR images is blurred,and detection and extraction are difficult,and target area extraction is a key problem in multi-level resolution imaging methods.Fast and accurate detection and extraction can improve the speed and efficiency of this method.In view of the fact that the current deep learning SAR image detection method has advantages over traditional methods in accuracy and computational efficiency,the target detection method based on deep learning is introduced into multi-level resolution SAR imaging detection,and it is found that the existing deep learning detection method is in In low-resolution SAR images,the detection accuracy is low,and the accuracy and speed are mutually restricted.To this end,this thesis proposes a lowresolution SAR image detection method based on knowledge distillation and probability distribution bounding box representation.Aiming at the problem of blurred target boundary in low-resolution images,the method replaces the Dirac function representation box used in the deep learning network with the probability distribution function representation box,and modifies the corresponding loss function according to the probability distribution representation,which significantly improves the detection accuracy.On this basis,in order to reduce the model parameters and improve the detection speed,a knowledge distillation mechanism is introduced in the process of locating the frame in the detection network,which greatly improves the detection accuracy and detection speed.In addition,this thesis also produces a low-resolution SAR image ship target dataset(LRSSDD)for the validation of the method.Finally,based on the LRSSDD data set,the proposed method is compared with the existing deep learning detection methods(Faster R-CNN,YOLO,SSD,FCOS,CenterNet).The results show that the accuracy(AP)of the proposed method reaches 49.2,which is excellent It is 48.5(Faster R-CNN),the highest among these methods,and the method in this thesis takes into account the detection speed(21FPS),which is higher than the 12FPS of Faster R-CNN.