Ship Detection Of SAR Images Based On Deep Learning

Synthetic Aperture Radar(SAR)ship detection can provide information support for military reconnaissance,civil detection and other tasks,which is an important part of SAR image interpretation.The result of SAR image ship detection can be used as a precondition for other interpretation tasks,and the detection result affects the performance of the entire SAR image automatic recognition system.In view of the problems encountered in SAR image ship detection and the advantages of deep learning.This thesis proposes a Ship Detection of SAR Images Based on Deep Learning.The neural network learns the abstract features of ship targets from image data,and learns nonlinear relationships by constructing multiple hidden layers and classifies and locates the ship target.Firstly,aiming at the problem of multi-scale,small target and strong interference of ship targets in SAR images,an enhancement method of feature fusion and feature denoising is proposed.The detection ability of the target detection algorithm for multi-scale and small target ships is improved by dense sampling,fusion of features between different levels and prediction of multi-scale feature map.By using the feature denoising module based on attention mechanism,the false detection rate of strong interference targets such as ports is greatly reduced and the accuracy of ship detection algorithm is improved.Secondly,in view of the dense ship arrangement in the port,the horizontal prediction box is prone to miss detection and loss of ship aspect ratio information,a two-stage rotating ship detection network is designed.By improving the classification regression subnetwork,the rotation Anchor setting and the matching method,the algorithm can predict the rotation detection box.The validity of the algorithm is verified by experiments.Finally,the ship detection problem of large-scale SAR images is solved by multi-GPU parallel method,which can complete the ship detection task of end-to-end SAR images in real time and accurately.