Ship Detection In SAR Image Based On Improved Faster R-CNN

With the gradual development of spaceborne radar,there are more and more satellites carrying high-resolution Synthetic Aperture Radar(Synthetic Aperture Radar),which makes high-resolution and ultra-high-resolution SAR image data more and more abundant.This makes the super-high resolution SAR image data more and more abundant,and in today’s ocean transport flow is gradually expanding,the marine surrounding environment is gradually complex,and large-scale ports and free trade zones are mixed,for the sea surface ship targets and ship targets in the port,SAR satellite has a broad application prospect in the case of a large range of target detection relying on the advantages of all day and all-weather.The application of SAR images to ship target detection has also become an important research direction in the field of remote sensing.In reality,the existing SAR image ship target detection algorithms mostly rely on manual intervention,which is low in efficiency and difficult to apply.Therefore,improving the automatic recognition ability of SAR image ship target detection algorithms has been a hot spot in recent years.In order to improve the interpretation efficiency of ship targets in SAR images and reduce manual intervention,this paper uses Faster R-CNN in the deep learning algorithm as the main framework to specifically improve SAR images.The main contents of this paper on SAR image ship target detection are as follows:1.By analyzing the characteristics of SAR imaging and comparing the advantages and disadvantages of SAR ship images with traditional optical images,SAR images have the advantages of all-time and all-weather compared with optical images;the analysis of the imaging mechanism of SAR images shows that It is easily affected by coherent speckle noise.Analyze the coherent speckle noise model and statistical characteristics of the SAR image,combined with the existing coherent speckle noise filtering method,preprocess the SAR ship data set,and compare it with the unprocessed detection results.The results show that the algorithm is good for the data.The accuracy of the test results collected for preprocessing has been improved;2.The original deep learning algorithm Faster R-CNN is aimed at traditional optical images,the target proportion is moderate,and the image information is rich.It has multiple bands of gray information,which is convenient for the extraction of feature extraction network,and it is directly applied to the SAR image ship target The test results are very unsatisfactory.The image information is simple,and there is no rich information data such as color.The original VGG16 feature extraction network has a shallow depth and cannot extract image information well.The improved algorithm deepens the feature extraction network and uses Res Net101 to strengthen the feature extraction capability.Experiments verify that the accuracy of the improved algorithm is significantly higher than that before the improvement;3.In the high-resolution spaceborne SAR image,the position of the spaceborne radar is very far away from the sea level,resulting in a relatively small proportion of the ship image in the SAR image.Aiming at the problem of a relatively small proportion of the SAR ship target,the algorithm area is recommended The network part is improved,and the K-means clustering algorithm is used to improve the value of the preset anchor point box,and the region of interest(Region of Interest Align,Ro I Align)that is more sensitive to small targets is introduced into the region suggestion network.For the above improvements,this article compares the algorithm before and after the improvement.The experimental results show that the accuracy of the improved algorithm has been improved.