SAR Image Target Recognition Based On Deep Convolutional Neural Network And Unsuperised K-means Feature

Image recognition is a representative application of pattern recognition,it processes,analyzes and understands the image via the computer to recognize different patterns.Along with the more and more important role which the Synthetic Aperture Radar has played in the military and commercial applications,a large amount of researchers begin to focus on the SAR image target recognition.The key technique in the SAR image recognition system is the Automatic Target Recognition(ATR).As the role that ATR has played in feature extracting and other aspects effectively,it become the domestic and foreign researching hot spot.The ATR devote itself to find the region of interest,extract the effective features from the target in the SAR image,and recognize the target through algorithm.Recent few years,a deal of algorithms have been proposed,like other algorithm of image recognition,it mainly includes three procedures: image pre-processing,feature extracting and target classification.The jobs of this paper are depend on the three steps and the jobs as following:(1)For SAR image pre-processing,as the SAR image has many background noise,so,firstly,this paper is focus on the searching of the Region of Interest(ROI),in a general way,the target of the image is in the center of the image,so centroid method is used to search the region of interest of the SAR image to make the background noise removed mostly.Then,in machine learning,especially the unsupervised learning and deep learning,the learning algorithm has a high demand of the scale of the database.For the SAR image,the scale of original SAR image is often small.So this paper proposed two methods of data augmentation to get more data.Furthermore,more training data is generated by using the data augmentation via rotating the azimuth of the target and adding a random integer to the original image.(2)Deep learning is another researching hot spot,which has many breakthroughs in many fields,such as image recognition,speech recognition and so on.Convolutional neural networks is one model of the deep learning algorithm.This paper trained a deep convolutional neural networks model,which named SARnet.SARnet has two convolutional layers,two pooling layers,two fully-connected layers,the model can learn high level features through the training procedure.The features of the testing data are extracted from the model,which istrained via the augmented database.The extracted features are used for the SAR ATR,and the result of the experiment is 95.68%,which is greater than other CNN models used in SAR image.(3)Much recent work in machine learning has focused on learning good features from unlabeled input data,For the feature extracting of the SAR image,firstly,we extract the ROI image,which size is 64×64,and then adopt the K-means unsupervised algorithm to learn the useful features through coding the patch of the image and adjusting the receptive field size of the patch through the augmented database,the model can learn more diverse features,Experimental results have shown that the proposed method can achieve the state-of-art accuracy,which is 96.67%.