SAR Ship Target Detection Method Based On CFAR And Two-stage Candidate Area Network

Synthetic Aperture Radar(SAR)has a strong penetrating ability,which can make it ignore the interference of weather factors and observe the ocean all day long.SAR-based ship target detection technology is one of the most important parts of maritime monitoring,and this technology is widely used in military and civilian fields.The SAR satellite GF-3 developed by my country can provide high-resolution SAR images of the ocean on a global scale,which provides strong data support for the ship detection work of SAR images.This paper mainly studies the ship target detection method on the SAR image of GF-3.The specific work is as follows:1.A complete process of CFAR detection is proposed for SAR images containing land areas.After using the improved OTSU algorithm to remove the land,use CFAR to detect the pure ocean area,and use the feature information such as aspect ratio and area to further remove the false targets that do not conform to the ship’s characteristics.Aiming at the slow running speed of CFAR,a fast CFAR detection algorithm is studied based on theG~0-distributed sea clutter statistical model.The two-dimensional Fourier transform is used to replace the traditional sliding window movement,and two local CFAR detections are used to make the threshold The calculation is more accurate.The experimental results show that the algorithm in this paper is faster than the traditional CFAR detection speed and has a lower missed detection rate.2.Aiming at the problem that the two-stage detection method Faster R-CNN has an unsatisfactory detection effect in SAR image detection,this paper proposes a SAR image ship target detection method based on the improved Cascade R-CNN network.improved.First,the CBAM(Convolutional Block Attention Module)attention module and the feature pyramid network are added to the feature extraction network,so that the network can fuse the features of different levels and help to extract the effective features of the target;Kmeans is used in the candidate region generation network to reset the ratio of the anchor point frame,so that the generated candidate frame is more suitable for the ship target;the three-stage cascaded detector is used in the detector module to improve the quality and quantity of positive samples.The experiment uses the public data set SSDD and expands it,and compares the algorithm in this paper with the three methods of Faster R-CNN,Yolov3 and Retina Net in terms of precision,recall and AP.The results show that the method in this paper is in three indicators.All of the above methods are better than the other three methods,which verifies the effectiveness of the network in this paper.3.For the wide SAR image of GF-3,this paper proposes a large-scale SAR image ship detection process,and the image is repeatedly cropped to prevent the cut ship from being segmented and lead to missed detection.Aiming at the problem that the detection network is slow to detect large-size images and there are many false alarm targets in the land area,a detection method combining CFAR and Cascade R-CNN is proposed.The sliced area of the target is sent to the subsequent detection network,and the sea surface area and pure land area that do not contain suspected targets are eliminated.The amount of data sent to the subsequent network is reduced,the detection speed is accelerated,and the false alarm targets in the pure land area can be eliminated to improve the detection accuracy.