SAR Image Recognition Based On CNN

Synthetic Aperture Radar(SAR)is a high resolution imaging radar and active microwave remote sensing detector.The first application of SAR was in the late 1950 s,when it was installed on RB-47 A and RB-57 D strategic reconnaissance aircraft.After decades of progress,SAR technology has been quite mature,each country has its own SAR development plan,various types of SAR have appeared in front of us,they play an important role in many areas.Then,with the rapid development of science and technology,the research of SAR is more and more comprehensive,especially in SAR recognition and other aspects.In this paper,two methods of SAR image recognition are proposed.The main work is as follows.:In this paper,Hu invariant moments,the Affine invariant moments and the Zernike invariant moments of SAR images are used to optimize SVM by adjusting PSO after adjusting learning factors,then SAR image classification and recognition algorithm based on improved PSO-SVM is proposed.This method mainly regulates the asynchronous learning factor of PSO,and strengthens the learning ability of particles.In terms of algorithm performance,it not only reduces the probability of particles falling into local optimum,but improves the convergence of algorithm effectively.Finally,the SAR image classification experiment is carried out.Secondly,the SAR image is recognized by using CNN-ELM algorithm.Because the features that are beneficial to classification can not be found in a short time,considering to use CNN.The images can be considered directly as the input of CNN,and the last layer of CNN can also be used as classifier directly.So CNN itself forms a classification and recognition system.In order to achieve better recognition effect,the activation function of traditional CNN is changed to Rule function,and SAR image features are obtained throughCNN.Finally,the feature is used as the input of the Extreme Learning Machine(ELM),so the CNN-ELM algorithm is obtained and used to recognize the SAR images.Finally,experiments are carried out on MSTAR public datasets of the United States,and the CNN classification and recognition effect is tested.The experimental results show that the algorithm not only realizes the sparsity of the network,alleviates the over-fitting problem,speeds up the convergence speed of the network,but also has a high recognition rate.