Target Recognition Of SAR Based On Convolutional Neural Network

Synthetic aperture radar(SAR)target recognition is an important part of SAR image interpretation.As a kind of high-resolution microwave imaging radar,it can automatically extract the characteristics of the target without the interference of light,climate and other factors.Since it was invented,it has played an important role in military reconnaissance and other fields.SAR image contains a lot of speckle noise.With the increasing amount of data,traditional methods can not solve the problem of image recognition.In additon,the research of deep learning is gradually becoming a hot spot in the field of machine learning,which contains special network structure with multiple hidden layers can automatically learn the characteristics of the image from the huge training data,thus greatly improving image recognition efficiency.In this paper,the problem of SAR target recognition is studied by combining convolutional neural network with a variety of classifiers.The main contents are as follows:(1)Only when the training samples are sufficient,CNN can learn representative features and prevent over fitting.So before the experiment,we need to use translation,rotation,image and other methods to expand the amount of original data.(2)Based on the traditional convolution neural network,an algorithm combining CNN,principal component analysis(PCA)and decision tree(DT)is proposed,which is recorded as CNN-PCA-DT: the feature vector of SAR image is extracted by convolution and pooling operation of CNN,and then the main feature is extracted by dimension reduction of PCA.Finally,DT classifier is used to replace the original softmax classifier in CNN to realize SAR target recognition,and the recognition accuracy is 99.60%,3.27% higher than the original network model.(3)On the basis of the traditional convolutional neural network,the softmax classifier is replaced by the random forest(RF),and the algorithm of combining CNN with the randomforest classifier is obtained,which is recorded as CNN-RF: in CNN-RF,the feature vector of SAR image extracted by CNN is input to the RF classifier,and the final recognition accuracy is 99.33%,3.0% higher than the original network model.Finally,the experimental results show that the proposed algorithm achieves a high recognition accuracy in MSTAR data set,and effectively improves the recognition accuracy of SAR image target compared with other methods,which shows the effectiveness of the experiment.