Abstract:Aiming at the problem of the difficulty of high-resolution synthetic aperture radar (SAR)image acquisition and poor feature characterization ability of low-resolution SAR image, thispaper proposes a method of an automatic target recognition method for SAR images based ona super-resolution generative adversarial network (SRGAN) and deep convolutional neural network(DCNN). First, the threshold segmentation is utilized to eliminate the SAR image background clutterand speckle noise and accurately extract target area of interest. Second, the low-resolution SARimage is enhanced through SRGAN to improve the visual resolution and the feature characterizationability of target in the SAR image. Third, the automatic classification and recognition for SARimage is realized by using DCNN with good generalization performance. Finally, the open data set,moving and stationary target acquisition and recognition, is utilized and good recognition results areobtained under standard operating condition and extended operating conditions, which verify theeffectiveness, robustness, and good generalization performance of the proposed method.Keywords:synthetic aperture radar (SAR); automatic target recognition (ATR); image segmentation;super-resolution generative adversarial network (SRGAN); deep convolutional neural network(DCNN)1. IntroductionDue to its advantages of all-day, all-weather, and strong penetrating capability, synthetic apertureradar (SAR) has been widely used in military and civil fields. SAR is a kind of active microwaveimaging radar, which can obtain two-dimensional (2-D) images with high resolution [1–4]. Theautomatic target recognition (ATR) is for SAR images to extract stable and iconic features based on SARimages, and determine its category attribute and confirm its particular copies of the same class, whichcan be applied to battlefield monitoring, guidance attack, attack effect assessment, marine resourcedetection, environmental geomorphology detection, and natural disaster assessment, and has vitalresearch significance. ATR also plays an important role in the electronic warfare (EW) and electronicintelligence (ELINT) systems [5,6]. The initial artificial interpretation for SAR images is inefficientand overly dependent on subjective factors. Therefore, in recent years, ATR for SAR images hasattracted significant attention from many experts, which is one of the most popular topics in currentresearch [7–9].Remote Sens.2019, 11, 135; doi:10.3390/rs11020135 www.mdpi.com/journal/remotesensingRemote Sens.2019, 11, 135 2 of 23The generalized ATR for SAR images can be divided into three levels: SAR target discrimination,SAR target classification, and SAR target recognition. SAR target discrimination can only distinguishthe difference between SAR targets. SAR target classification predicts the class of a target in the SARimage on the basis of SAR target discrimination. SAR target recognition confirms the specific copiesof the same class of targets in SAR images based on target discrimination and target classification.Generally, when we say the target recognition is a narrow sense of target recognition, it only meansthe highest level of target recognition. This paper mainly identifies target recognition. It mainlyincludes three steps: target detection, discrimination, and recognition [8]. Target detection extractsthe target region of interest from a SAR image using image segmentation technique to eliminatebackground clutter and speckle noise, enhance the target region, and weaken the influence of thebackground on recognition. The process of target discrimination is mainly the process of featureextraction, which extracts and integrates effective information in the SAR image and transforms theimage data into feature vectors. Good features have good intra-class aggregation and inter-classdifference in the classification space. Pei et al. [10] extracted SAR image features using 2-D principlecomponent analysis-based 2-D neighborhood virtual points discriminant embedding for SAR ATR.However, when new samples come in, features and models need to be relearned. This method’suniversality is low and it is time consuming. To overcome this problem, Dang et al. [11] used theincremental non-negative matrix decomposition method to study the features online to improve thecomputational efficiency and the universality of the model. After feature extraction, different classifierscan be designed to classify targets for SAR images. There are three mainstream paradigms of ATR forSAR images: template matching, model-based methods, and machine learning. Template matching isthe most common and typical one, which stores the physical features, structural feature, etc., extractedfrom the training samples in the template data set, and matches the target features of all samplesin the template library until matching rules are met to determine the information of the target to betested [8,12]. However, this method requires a large amount of computation and prior information. Theextracted features need to be manually designed, and it is difficult to fully explore the mutual relationsamong the massive amount of data. The basic idea behind the model-based classification method is toreplace the target feature templates stored in the target data set with solid model or scattering centermodel, which could construct a feature template in real time for recognition according to the specificconditions such as target posture. Verly et al. [13] achieved recognition results by extracting the length,area, location, and other features in control and matching them with the model library. However, thismethod needs to build the attribute diagram of target size, shape, etc., which is difficult to implement,and only applicable to specific scenarios.With the rapid development of computer hardware devices, machine learning is widely used inoptical image processing [14], speech recognition [15], speech separation [16], etc. In recent years, ATRmethods for SAR images based on machine learning have been widely used and achieved very goodresults. Verly et al. and Zhao et al. classified and recognized the ground vehicles, whose data wasfrom the moving and stationary target acquisition and recognition (MSTAR) [17], by using AdaBoostand a support vector machine based on a maximized classification boundary [13,18]. However, thesemethods require hand-designed features and empirical information, is heavily dependent on subjectivefactors, and had low universality. Wang et al. utilized the wavelet scattering network to extract waveletscattering coefficients as features [19]. Although the convolutional network was utilized, it also belongsto the traditional methods which contain three steps: feature extraction by hands, dimension reduction,and classification using different classifier. He et al. [20] utilized convolutional neural network (CNN)to classify SAR images, with a final recognition rate of 99.47%, but only seven categories of targets inthe MSTAR data set were classified. With the increase of layers, more and more parameters need to betrained for CNN. Meanwhile, overfitting is occurred easily, which leads to the network’s inability toconverge or to converge to the global optimum. To reduce the number of the network parameters, Chenet al. [21] proposed a SAR image target recognition method based on A-ConvNets, which removed allthe fully connected layers and only contained sparse connection layers. Asoftmaxactivation functionRemote Sens.2019, 11, 135 3 of 23was utilized at the end of the network to achieve the final classification. This method was verifiedusing MSTAR data set, and the recognition rate got 99%, which was higher than the traditional method.However, the recognition rate of this method for SAR images after segmentation is only 95.04%.Schumacher et al. [22,23] pointed out that the radar echo of each type of target in MSTAR data setcan only be recorded under a specific background, that is, there is a one-to-one relationship betweenthe target and the background, and the background can also be used as a feature of the target forclassification and recognition. Based on this, Zhou et al. [24] used the traditional CNN to classify theSAR image background in MSTAR, and obtained the recognition rate of 30–40%, which proved thatSAR image background can improve the recognition rate. At the same time, Zhou et al. proposeda large-marginsoftmax(LM-softmax) batch-normalization CNN (LM-BN-CNN) method, which hada better recognition rates under both standard operating condition (SOC) and extended operatingconditions (EOCs).However, if the SAR image quality is not good and the resolution is low, it will greatly affectthe correct recognition rate of SAR targets. The above methods are all based on the original SARimages, and the image quality is not improved and enhanced. In recent years, some researchershave done a lot of studies on image super-resolution reconstruction [25,26]. Image super-resolutionreconstruction techniques overcome the disadvantages of imaging equipment’s inherent resolution,breaks the limitation of imaging environment, and can obtain high-quality images, which is higherthan the physical resolution of the existing imaging system, at the lowest cost. The existingsuper-resolution reconstruction technique of a single frame image is mainly divided into three types:an interpolation-based method, reconstruction-based method, and learning-based method. With thehelp of machine learning techniques, the high frequency information loss of the low-resolution SARimage is estimated by learning the mapping relationship between low-resolution and high-resolutionSAR images in order to obtain the detailed information on the clear target, such as edge, contour,texture, etc. Thus, the image features characterization ability is enhanced, and the SAR image correctclassification coefficient is improved in this paper. Liu et al. [27] adopted a joint-learning-basedstrategy, combined with the characteristics of SAR image, to reconstruct a high-resolution SAR imagefrom low-resolution SAR image to achieve the global minimum of the super-resolution error andreduce speckle noise. Li et al. [28] utilized a Markov random field and Shearlet transformationto recover a super-resolution SAR image. The result of this method is better than the traditionalmethod, but the detailed texture information of the reconstructed image is still different from theoriginal image in visual effect. The super-resolution reconstruction method based on deep learninguses multi-layer neural network to directly establish the end-to-end nonlinear mapping relationshipbetween low-resolution and high-resolution images. Dong et al. [29] proposed a nonlinear regressionsuper-resolution reconstruction method using CNN, but this method has fewer layers and a smallerreceptive field. To overcome this problem, Kim et al. [30] achieved better results based on recursiveneural network super-resolution technology by adding the number of convolutional layers andreducing the number of network parameters. In recent years, the generative adversarial network(GAN) has been developing rapidly with its unique advantages. It uses the game confrontationprocess of a generator and a discriminator to realize new image formation [31,32]. The MSTARdata set was utilized as training set to generate more realistic SAR images by GAN to expand theSAR image data set. Leding et al. [33] improved GAN to obtain super-resolution GAN (SRGAN)by replacing the loss function based on mean square error (MSE) with the the loss function of thefeature map of visual geometry group (VGG) network. Under the condition of high magnification, thereconstruction of optical image from low-resolution to high-resolution was realized, and better visual