Ground Target Recognition Investigation In SAR Images Based On Deep Learning Network

Synthetic Aperture Radar(SAR)has been widely used in civil and military fields due to its great ability to work day and night under all weather conditions.SAR can achieve high-resolution imaging of Earth observation.After more than half a century of research and development,its theoretical research and imaging technology have been relatively mature.SAR Automatic Target Recognition(ATR)technology is dedicated to effectively detecting targets from complex terrain scenes and identifying the detected targets.Specifically,the ATR process is based on first the detection and localization of the target by the radar,then extracts the characteristics of the target according to target detection and orientation results from the radar echoes of the interested target and the background,and finally determines the attribute,category or model of the target of interest.The research of SAR ATR technology plays an important role in improving the military's command automation level,battlefield perception ability,offensive and defensive ability,national air defense and anti-missile capability and strategic early warning capability.It is a hot t opic of scholars at home and abroad.SAR imaging has some advantages over optical imaging,like its low sensitivity to weather conditions and its penetration ability through obstacles.However,the pixel brightness values and geometry data used to identify SAR images are susceptible to satellite sensors and propagation,which makes it difficult to detect and classify targets directly based on SAR images.In recent years,some SAR-ATR techniques have been able to alleviate this problem to some extent,such as Non-Negative Matrix Factorization(NMF),Principal Component Analysis(PCA)and Convolutional Neural Network(CNN),etc...,but the actual effect is poor and the adaptability is not high.In this context,this research is in-depth study of SAR-ATR-related theories and techniques based on deep learning networks.The main research content of the thesis is as follows:In order to study the effect of pre-processing on the SAR images recognition task,a Sparse Auto-Encoder(SAE)-CNN-Recognizer(SCR)is introduced into application of SAR ATR.Specifically,a two-step algorithm,the first step is using SAE as new way of raw data pre-processing to enhance the quality of SAR images,followed by a two-stage convolutional network architecture which is designed to automatically learn features from restored SAR images and classify them into different classes.Aiming at the problem of fusion the SAR image CNN features extracted from the fully connected layers,the paper proposes a new framework for SAR target recognition based on CNN of and Stacked Auto Encoder(SAE).The framework firstly extracts features from the fully connected layers of the two CNN architectures,and then using an effective feature combination method the extracted features are combined and fed into a SAE,which accurately solves the recognition problem of the combined features.The obtained results prove the effectiveness of the method.Afterwords,in order to study the effect of combination and fusion on the recognition accuracy of SAR iamges,a new framework,which is drawn from the most contemporary techniques in image processing and deep learning is introduced.It is a combination of three different CNNs.The three CNNs have the same architecture;the main difference is the sizes of the convolut ion and pooling kernel in each one of them: Coarse Grain CNN(CG-CNN),Middle Grain CNN(MG-CNN)and Fine Grain CNN(FG-CNN).Each CNN is trained from scratch using the chosen sizes of convolutional and pooling kernel.We investigate how to obtain better accuracy classification compared to the state of art results that use the same dataset without data augmentation.Capsule Networks(Caps Net)is an emerging deep learning network in the past three years,getting better performance for smaller sample collection.The paper applies the Caps Net algorithm to SAR ATR,which improves the accuracy of the current recognition problem by using two main techniques which are the feature routing by agreement and the vector-output capsules rather than scalar output feature detectors.The experimental results on the public MSTAR data set show that this method can successfully overcome the deficiencies of CNN in classifying SAR images.In order to further improve the recognition accuracy of CNN and Caps Net,the paper proposes a dual convolutional Caps Net method.This method uses the features extracted from the layers of the CNN network,and then uses Caps Net for classification and recognition,so that the two methods are combined.Compared with a single traditional CNN and Caps Net,this algorithm has better recognition accuracy.