Research On SAR Image Ship Detection Algorithm Based On Convolutional Neural Network

In recent years,with the rapid development of remote sensing technology,many scholars have devoted themselves to remote sensing research,making remote sensing a research hotspot with new theories and methods emerging one after another.With the launch of GF-3 satellites,domestic researchers have set up a new wave of research on Synthetic Aperture Radar(SAR).SAR uses microwave imaging to monitor the ground all-day long and all-whether without being affected by ground cloud,light,wind and rain.It can availably identify camouflage and penetrate the cover to obtain ground information.In the civil field,SAR can effectively monitor mountains,forests,oceans,natural resources and natural disasters.In the military field,SAR can provide reconnaissance,obtain battlefield information and enhance battlefield advantages.Ship detection is an important research field of SAR.The imaging mechanism of SAR makes SAR images full of coherent speckle noise,which increases the difficulty of traditional ship detection.Therefore,it is of great significance to find a method for SAR image target detection with better detection results and better universality.In this paper,an improved ship detection algorithm for R-FCN SAR image is proposed on the basis of convolutional neural network.Among traditional ship detection algorithms,Constant False Alarm Rate(CFAR)detection algorithm is the most widely used and studied algorithm.CFAR realizes detection by fully fitting background clutter and utilizing the difference between background clutter and target.However,the background clutter distribution of SAR image is complex,and the clutter model is difficult to fit background clutter in various states.Different from the traditional method,this paper introduces the convolutional neural network into the SAR image ship detection,and identifies the ship by learning the ship characteristics of different forms,which avoids the SAR image preprocessing and the complex background clutter fitting,and has a higher detection effect in the complex environment.This paper mainly completes the following research contents:First,construct the ship data sets.Convolutional neural network training requires a considerable number of data sets,but there is no publicly available ship data sets.In order to obtain enough ship samples,this paper selects the open Sentinel-1 satellite data and makes use of the ship samples uploaded by experts to make the data sets.Second,introduce the Region-based Fully Convolutional Networks(R-FCN).Because the imaging principle of SAR image is different from that of optical image,the detection effect of original R-FCN in SAR image is not satisfactory.In this paper,according to the characteristics of SAR images,mixed scale convolution kernel processing is performed on ResNet of R-FCN,which enables the feature extraction network to suppress the influence of speckle noise and effectively extract ship features.The low-resolution Sentinel-1 satellite data is selected to test the algorithm in this paper,and the detection accuracy is as high as 97.46%,significantly higher than the original R-FCN detection algorithm,proving the effectiveness of the proposed algorithm.At the same time,high-resolution GF-3 satellite SAR images are selected to test the algorithm,and the detection accuracy reaches as high as 97.37%,proving the universality of the algorithm.