1.Schoolof Electronic and Information Engineering,Beihang University,Beijing 100191,China;

2.Space Mechatronic Systems Technology Laboratory,Departmentof Design,Manufacture and Engineering Management,University of Strathclyde,Glasgow G1 1XJ,UK;

3.Cognitive Signal-Image and Control Processing Research Laboratory,Schoolof Natural Sciences, University of Stirling,Stirling FK9 4LA,UK

Targetdetection and recognition in SAR imagery based on KFDA

FeiGao1,Jingyuan Mei1,Jinping Sun1,*,Jun Wang1,Erfu Yang2,and Amir Hussain3

1.Schoolof Electronic and Information Engineering,Beihang University,Beijing 100191,China;

2.Space Mechatronic Systems Technology Laboratory,Departmentof Design,Manufacture and Engineering Management,University of Strathclyde,Glasgow G1 1XJ,UK;

3.Cognitive Signal-Image and Control Processing Research Laboratory,Schoolof Natural Sciences, University of Stirling,Stirling FK9 4LA,UK

Current research on target detection and recognition from synthetic aperture radar(SAR)images is usually carried out separately.It is dif fi cult to verify the ability of a target recognition algorithm for adapting to changes in the environment.To realize the whole process of SAR automatic targetrecognition(ATR),especially for the detection and recognition of vehicles,an algorithm based on kernel fi sher discriminant analysis(KFDA)is proposed. First,in order to make a better description of the difference between the background and the target,KFDA is extended to the detection part.Image samples are obtained with a dual-window approach and features ofthe inner and outer window samples are extracted by using KFDA.The difference between the features of inner and outer window samples is compared with a threshold to determine whether a vehicle exists.Second,for the target area, we propose an improved KFDA-IMED(image Euclidean distance) combined with a support vector machine(SVM)to recognize the vehicles.Experimental results validate the performance of our method.On the detection task,our proposed method obtains not only a high detection rate but also a low false alarm rate without using any prior information.For the recognition task,our method overcomes the SAR image aspect angle sensitivity,reduces the requirements for image preprocessing and improves the recognition rate.

synthetic aperture radar(SAR),target detection,kernel fi sher discriminant analysis(KFDA),target recognition,image Euclidean distance(IMED),supportvector machine(SVM).

1.Introduction

Synthetic aperture radar(SAR)extends the radar signalto two dimensions and greatly enriches the information acquired.It has become a landmark in the history of developmentof radar technology.

With the increase of radar coverage areas and continuous development in SAR imaging technology,the amount of SAR images has far exceeded the limit of current image interpretation capability.The development of SAR sensors is moving to high-resolution,multi-mode, multi-frequency,multi-polarization and multi-band,which makes the information in SAR images more abundantbut also demands higher requirements for targetdetection and recognition.Thus SARautomatic targetrecognition(ATR) has become one of the hottesttopics in the fi eld of military defense over the past20 years.In particular,vehicle(such as armored vehicle,tank and car)detection and recognition are two importantbranchesof SARATR.Researchers have thus proposed a variety of effective algorithms for these two emerging areas.

There are two categories of algorithms for vehicle detection in SAR images.The fi rst is based on the contrast of pixels such as constant false alarm rate(CFAR)which is the most classic and the most widely used algorithm. Novak etal.[1]derived a highly in fl uentialtwo-parameter CFAR detector with an assumption thatthe amplitude distribution of the clutter is Gaussian,which opened a new chapter for target detection in SAR images.The developmentof many derivative algorithms such as ordered statistic(OS)-CFAR,variability index(VI)-CFAR and extended fractal(EF)-CFAR has since followed.However,in the current situation,CFAR also faces enormous challenges. The main problem is that CFAR always yields a high false alarm rate under a complex environmentand weak-targetcondition as a resultof ignoring the target's own features. The detection result also strongly depends on the selected distribution modelwhile there are considerable dif fi culties in building a universaldistribution model.The second category is based on a numberofotherfeatures ofSARimages. Researchers[2,3]introduced the conceptofknowledge and proposed a detection method based on prior knowledge and contextinformation respectively.These algorithms can fully use the prior information although much of it is often hard to obtain for strange geographicalenvironments. Researchers in[4]developed a multi-resolution algorithm for detecting man-made objects in SAR images.Authors [5]used shadow information as a prescreening feature to improve detection performance.One of the disadvantages for works reported in[4,5]is that they apply complex advanced features such as multi-resolution and shadow.The access to these advanced features is not easy and also a large amount of calculation is needed.Recently,a novel vehicle detection algorithm based on the visual attention mechanism[6]has become a major research focus and is attracting growing attention of researchers.

There are also two categories of vehicle recognition algorithms in SAR images.The fi rst is the model-based algorithm.In the literature[7-9]Bhanu et al.described their contribution to model building.They predicted a series of associated candidate targets through the mathematical models and made a judgment by matching.The disadvantage of such algorithms is that building an SAR image model needs a high level of theoretical basis and calculation.The second category is the template-based algorithm.Previous researchers[10]proposed a typical method using original SAR images or sub-images as templates.However,this method is extremely sensitive to changes in depression angle and aspectangle so that it requires a large number of templates.In recent years with the rapid developmentoffeature extraction algorithms and classi fi cation algorithms,many feature-based classi fi cation algorithms for SAR images are proposed.For feature extraction algorithms,works reported in[11,12]featured scattering centers and used the Prony modeland the attributed model respectively.The work of[13]demonstrated the feasibility ofrecognizing vehicle targets in SAR images using the peak feature.Work in[14]presented twenty features(namely the standard deviation feature, fractal dimension feature,etc.)as an evolutionary feature synthesis.The selection of these features is a dif fi culttask. If we select only one feature,this re fl ects the unilateral information of the targets,which may cause large deviation with a small number of training samples.If some of the features are combined as feature synthesis,different syntheses lead to different recognition results since each feature may not be valid in a certain condition.Complexity can be increased by using more features,yet it does not necessarily mean a better result.However,this dif ficult task gives rise to various linear or nonlinear transformation algorithms,which are another importantbranch of feature extraction algorithms.These transformation algorithms try to extractglobalfeatures from the entire image, which can remove redundant information and achieve dimensionality reduction.Linear principle componentanalysis(PCA)is applied in[15]to extract features.The disadvantage of PCA is that it only considers second-order correlation between the samples and ignores the higher order statistics.In other works[16],classic nonlinear kernel principalcomponentanalysis(KPCA)is used butsimulation results show that the ability of KPCA to extract discriminative features is not effective.In recent years,the idea of manifold ways of perception[17]has also been introduced to the fi eld of SAR image feature extraction and become a popular research focus.For classi fi cation algorithms following the feature extraction stage,researchers [18,19]used supportvector machine(SVM)for SAR vehicle recognition and set off a new wave of research.Adaboostis also used in[20],which has also been widely promoted.Currently SVMand Adaboosthave demonstrated a lotbetter performance compared with other algorithms.

A vehicle is often presented as a substantially rectangular bright spot in SAR images.With the improvement of camou fl age technology,strong scatters of the vehicle are reduced so that target pixels and background pixels have less difference.When the vehicle is planning to escape detection and remains adjacent to other structures in the complicated environment,key pixels are consequently decreasing and interference pixels increasing. Some importantfeatures like scattering centers,objectoutline and shadow information are destroyed,so that traditional detection algorithms based on contrast such as CFAR and others based on feature synthesis do notdeliver the required performance.

In this paper,we attemptto use the global features difference between the vehicle and the surrounding environment to improve the detection performance of hidden targets.As kernelmethods have been successfully applied to classi fi cation,we hope kernel methods can be introduced into detection in SAR images,which is essentially a classi fi cation problem between two classes.Kernel fi sher discriminantanalysis(KFDA)is a well-known discriminative feature extraction criterion of the pattern class.Itlooks for a projection direction in a high-dimensionalspace,which can make the samples of differentclasses separated as far as possible with samples in the same class gathered as close as possible.Thus,after acquiring the target sam-ples and clutter samples with a dual-window approach, the use of KFDA to calculate the difference between the two classes to achieve target detection may exhibit certain advantages.Simulation results show that application of the KFDA to SAR vehicle detection demonstrates performance.Itconsiders targetfeatures so that it can reduce false alarms which are highly brightand similar to the targets in traditionaltarget detection algorithms[1].Furthermore,without the use of prior information,as in literature[2,3],our KFDA method also avoids the dif fi culty of selecting and extracting advanced features[4,5]and obtains a high detection rate and low false alarm rate.For SAR targetrecognition,the differences between each class of vehicle are subtle and highly localized in few pixels. And forthe case ofreal-world conditions,vehicles include variations in aspect angle,target con fi gurations and obscuration due to occlusion and layover.These dif fi culties place great burden on algorithms using one feature or a feature synthesis.In this paper,we attempt to exploit the ef fi ciency of KFDA for extracting global nonlinear features of SAR vehicles in order to improve the recognition rate.In order to clearly re fl ect the spatial relationship and the deformation degree between two images,image Euclidean distance(IMED)is embedded in KFDA as an improved KFDA-IMED.Simulation results demonstrate that our KFDA-IMED method produces good recognition results for both the exposed targets and hidden targets and avoids the dif fi culties of selecting features as in literature [11-14]withoutpre-processing and image segmentation.

This paper proposes an organically integrated vehicle detection and recognition algorithm for SARimages based on KFDA.If training samples are not available,detection can still be achieved.Or if training samples are available, it can also enable vehicle recognition and validate the detection results in auxiliary.Furthermore,clutter samples can be collected from some of the background regions in the recognition partin order to reduce the number of false alarms generated in the detection part.

This paper is structured as follows.Section 2 presents KFDA(KFDA-IMED).The proposed SAR vehicle detection and recognition algorithm based on KFDA is described in Section 3.In Section 4,the experimentalresults of SAR vehicle detection and recognition are presented. We also analyze the adaptability of the recognition algorithm to the diversity of the environment clutter and the in fl uence of size of target image chip on the recognition rate.Section 5 presents some concluding remarks.

2.KFDAand improved KFDA-IMED

Assuming two different classes of n dimensional input samples can be projected to a higher dimensional feature space F by a nonlinear mappingφ:x∈R→φ(x)∈F,better linear separability of the samples can be obtained in the higher dimensional feature space.Following this their features can be extracted through the well-known Fisher's linear discriminantanalysis(FLDA)[21].

Let X denote an input sample set comprising N samples;Xidenote a subsetof X comprising Nisamples;and c(c?n)denote the number of classes,thus we haveThe input samplesand xj∈X can be mapped respectively in the higher dimensionalfeature space F The within-class scatter matrix of samples in the space F is de fi ned byis the mean of each class of samples in the space F andcharacterizes the degree of aggregation of internal samples in each class.While the between-class scatter matrix of samples in the space F is de fi ned byis the mean of

all the samples in the space F andcharacterizes the degree of dispersion among the differentclasses.

In the space F,FLDAlooks forthe projection direction w to make the samples of differentclasses separate as far as possible,whilst the samples in the same class are gathered as close as possible.This goalcan be written as

which is termed the so-called Rayleigh quotient[22].

The derivation above is the Fisher's criterion in the highdimensionalfeature space.In orderto derive KFDA,based on reproducing kerneltheory,w can be expanded as

whereαjis the coef fi cient.With(4),(3)can be changed to

For every input sample x,the features extracted by KFDA is actually the projection ofφ(x)in the optimaldirection w.Thus,each dimensionalelementof the features is

When the radial basis function(RBF)k(x,y)=is selected as the kernel function k(x,y),as x and y are two images not common vectors in targetrecognition,it only calculates the traditional Euclidean distance between them and ignores the spatialrelationships among the pixels so itis extremely sensitive even to smalldeformations.To tackle this disadvantage,we embed IMED in KFDA[24].IMEDcan re fl ectthe correlation among the image samples better since itis invariantto the linear transformation of the images,and depends on the extentof the deformation.

Let P1denote two H by L images,where H×L= n.The element gij(i,j=1,2,...,n)of the image metric matrix Gn×nre fl ects the relationship of position between pixel Piand pixel Pj.Assuming the location of Piis(h,l) and the location of Pjis(h?,l?),gijcan be written asThus,IMED between P1and P2is

Then the calculation of RBF with IMED is

For large images,the evaluation of G is expensive.The standardizing transform(ST)[24]or Kronecker product [25]can be introduced to simplify the calculations.

In the above derivation,the process of calculating Rayleigh quotientand extracting the features are based on the assumption that the within-class scatter matrix Kwis a nonsingular matrix.However,according to Fisher's criterion,Kwis a nonsingular matrix only if the number of training samples is greater than the dimension of the sample space.Otherwise Kwis a singular matrix and thus the inverse does not exist.To avoid such ill-conditioned settings,Kwcan be repalced by Kw≈Kw+κI.Here I is an identity matrix of the same order of Kw,and the constantκwhich acts as a disturbance,is small and greater than zero.The properties of Kw≈Kw+κI willbe dominated byκifκis too large,which means the degree ofaggregation of the internalsamples in each class is restricted. This willmake the feature eigenvectorαdisturbed,and the features ofeach class relatively dispersed,and the recognition rate gets reduced.κis often setto beκ?10?3.

KFDA and improved KFDA-IMEDabove constitute the core theory of the proposed algorithms in this paper.Their fl ow chartis shown in Fig.1.

Fig.1 Flow-chart of KFDA and improved KFDA-IMED

3.Algorithm description

3.1 Vehicle detection

The SAR vehicle detection algorithm proposed in this paper introduces KFDA to extractthe discriminative features of the vehicle and its surrounding environmentand calculates the difference between features to determine its existence.An edge preserving smoothing operation and grey scale quantization[2]are performed on the original SAR image in order to reduce the impactof speckle noise while preserving the boundary edges of a vehicle.

In general,a target has stronger scattering properties than its surrounding area while the target pixels are only a tiny part of the whole SAR image.Therefore we can choose an appropriate global threshold to achieve image binarization and determine possible targetpixels so we do notneed to carry outtraversalsearch forthe overallimage. When the vehicle stays adjacentto trees forhiding,another advantage for the global threshold operation is that it can separate the pixels of trees and the pixels of vehicles.For the case of trees,there are many gaps between leaves,and partof the electromagnetic waves willbe re fl ected directly by trees while some will be re fl ected by the ground.From Fig.2 we can see thattrees show a mixture of brightspots and dark spots.Fig.3 shows the results of performing the edge preserving smoothing operation,grey scale quantization and the global threshold operation.It is obvious that the trees are separated into discrete points by the global threshold so that it is conducive to detect vehicles around the trees.However in Fig.3,some brightspots can stillbe found,which are similar to the vehicles and produced by trees.False alarms may be caused by these brightspots.In subsequentprocessing,we willtry to eliminate these kinds of bright spots through the extraction of appropriate features of KFDA.

After the global threshold operation,two obvious problems stillconfrontus.First,there are stilla greatnumberof discrete non-targetpoints in the image.Second,some dark pointsofthe vehicles are excluded.Morphological fi ltering can be used to eliminate those discrete and smallpoints and make the targetregion more homogeneous,whilstpreserving its edge information[26].

Fig.2 Original SAR image

All the left bright spots in the binary image can belong to the vehicles,which need to be examined one by one.Before this,we perform image segmentation with connected component analysis,wherein each bright spot is assigned an index number and the coordinate of each bright spotpixelis recorded.This way,traversing all the pixels in the subsequentdetection process is notneeded.

Fig.3 SAR image after the globalthreshold operation

Until now,the number of bright spots in the binary image may still be large.While some obvious non-target spots can be simply excluded on the basis of their size or shape.For example,a long and thin brightspot can be ignored directly because its shape does not match the shape of a vehicle.

Next,the KFDA is operated with the dual-window approach which is the core steps of our proposed detection algorithm.The size and shape of the inner-window is determined according to the generalsize and shape of vehicles.The shape of the protecting window and the outerwindow is the same as the inner-window.The protecting window is slightly largerthan the inner-window whilststill smaller than the outer-window.Finally,the numberofpixels in the outer-window is the same as in the inner-window. The dual-window approach used in this paper is shown in Fig.4.The center of the dual-window can be located in each bright spot's geometric center,which is calculated with the coordinates of pixels of each bright spot in the originalSAR image,using the index number.

Fig.4 Form ofdual-window

In detection,we make the current inner-window move to differentdirections in the protecting window to geta series of inner-window samples,which are termed the innerwindow sample set.Meanwhile,the currentouter-window is moved to the same directions as the inner-window to get a series ofouter-window samples termed the outer-window sample set.Since the outer-window has a certain width, some partof the targetpixels may also be covered.Therefore,we set the number of leaked pixels to be D,which means that the D brightest pixels in each outer-window sample and the D darkest pixels in each inner-window sample will be eliminated.The value of D can be determined based on the width of the outer-window.

The complex diversity of the SAR image makes it necessary to normalize the amplitude of each inner-window sample and each outer-window sample.The normalization formula is given by

where x is the vector representation of each sample and xNormalizedthe vector representation of the normalized x.

Let X1denote the normalized inner-window sample set and X2the normalized outer-window sample set.The feature of currentinner-window sample ziand the feature of currentouter-window sample zoare extracted with KFDA through equations(1)-(7).Both ziand zoare one dimensional.The detection is achieved by comparing the difference between ziand zowith a threshold as follows:

where H1implies thatthe currentinner-window sample is a vehicle,H0implies the current inner-window sample is clutter,and T is the threshold.

The above SAR vehicle detection algorithm based on KFDA is outlined in the upper half of Fig.5,which shows that the detection result can be outputdirectly.The lower half of Fig.5 also shows our proposed recognition algorithm based on the KFDA-IMED when training samples are available.

3.2 Vehicle recognition

After the detection stage,the types of H1image samples can be recognized if we have the same types of vehicles in our training sample database.In this case,fi rstly,the H1image chips are collected from the original SAR image. Second,the amplitude of each training sample and each H1image chip should be normalized with(11).Third,the features of training samples and H1image chips are extracted with the KFDA-IMED through(1)-(10).The features of each training sample or each H1image chip can be represented as a c?1 dimensionalvector.

In this paper,the well-known SVM is chosen for features classi fi cation.The reasons are two-fold:fi rstly,SVM is based on the structural risk minimization basis so ensures theoreticalgeneralization;secondly,the SVMhas advantages in solving smallsample size and nonlinear problems and is relatively insensitive to the representation of features.SVM is trained with the features of training samples.The recognition of H1image chips is executed according to the“one-against-one”multiclass strategy.

Fig.5 Flow-chart of SAR vehicle detection and recognition algorithm based on KFDA

In the following step,we address a problem when false alarms in the detection may be incorrectly recognized as one type of vehicle.For this case,some environmental samples can be collected before recognition in certain areas which are around yet do not cover the H1image chips.We hope this approach can make the false alarms recognizable as environmentsamplesand thus be removed. However,this operation also raises anotherproblem.Ifthe undetected targets are collected as environment samples, the credibility of the recognition results is seriously affected.Hence,it is necessary to setthe threshold T lower to guarantee the absence or the presence of only few undetected targets before the environmentsamples collection step.

In much ofthe literature[15,16,26,27],researchers have performed some pre-processing operations for enhancing recognition,such as log transformation,Fourier transform,image segmentation,power-law transformation,image fi ltering and so on,and consequently obtained a high recognition rate.In this paper,however,after obtaining H1image chips from the original SAR image,only the simple normalization operation is applied.The main reasons can be summarized as follows:fi rst,we believe the KFDA-IMED demonstrates a consistently high ef fi ciency in terms of feature extraction and can reduce the requirements forimage preprocessing;second,these complex preprocessing operations do not lead to a de fi nitively higher recognition rate forour KFDA-IMED.Speci fi cally,ourexperience has shown thatnotevery pre-processing operation has a positive effect on the recognition rate for different feature extractions.For the case of our proposed KFDAIMED method,the recognition rate was not found to be improved by any of the otherpre-processing operations reported in the literature.

4.Experiments

4.1 Machine learning for the vehicle recognition experiment

In this experiment,spotlight SAR images of ground vehicles in the moving and stationary target acquisition and recognition(MSTAR)database[28]is used.The database provides many different types of vehicle samples which can be used for quantitatively analyzing the training process and recognition results of the algorithm proposed in this paper.The original size of SAR vehicle sample is 128×128 pixels.The resolution is 0.3 m×0.3 m.The azimuth coverage of each type of the vehicle is 0°to 360°and the intervalis approximately 1°.We choose three distincttypes of these in the database,namely:three BMP2 s (sn-c21,sn-9563,sn-9566),one BTR70(sn-c71)and three T72 s(sn-132,sn-812,sn-s7).The different serials in the same type are mainly because of the con fi guration variants [29],which are termed as varianttargets.

The training set includes three training sample sets(BMP2sn-c21,BTR70sn-c71 and T72sn-132)at 17 depression angle.The number of samples in each training set is 232.The testing set includes all seven serials at 15°depression angle.BMP2sn-c21,BTR70sn-c71 and T72sn-132 are the same serialtargetswhile BMP2sn-9563, BMP2sn-9566,T72sn-812 and T72sn-s7 are the variant targets.The number of samples in each testing setis 191.

Five recognition algorithms are compared in this experiment:SVM,KPCA+SVM,KPCA+FLDA+SVM, KFDA+SVM,KFDA-IMED+SVM.The kernel functions in KPCA and KFDA are selected as the easier-to-control polynomialfunction k(x,y)=(x·y+1)l,where l=5. The kernel function in KFDA-IMED is selected as the improved RBF k(x,y)whereσis the mean of IMED of all the training samples. The constantκis 10?4.The kernelfunction in SVMis selected as the RBF k(x,y)=exp(?0.6||x?y||2).The obtained recognition rates are listed in Table 1.

Table 1 Recognition rates of five different algorithms%

As shown in Table 1,inputting the normalized samples into SVM directly produces good results,which demonstrates that the SVM can effectively overcome the SAR image aspect angle sensitivity problem and is suitable for SAR image processing.However,the number of normalized sample dimensions is 16 384(128×128),which contains a lot of redundant information and leads to computational complexity problems.If the KPCA is used to extract features,the number of dimensions can be effectively reduced from 16 384 to 696.The KPCA approach facilitates calculations and its recognition rate of identical serial targets increases,while the rate of variant targets falls.This demonstrates that the KPCA is not ef ficient in extracting discriminative features of variant targets.For the case of KPCA+FLDA+SVM,the recognition rates of same serialtargets and varianttargets are both better than the KPCA+SVM and SVM methods.Meanwhile, the number of sample dimensions is compressed into two, which demonstrates thatthe introduction of the FLDA can improve the ability to extract discriminative features.For KFDA+SVM,both rates are higher than 90%.Introducing IMED on this basis,the recognition results are further improved,which proves the effectiveness of considering the spatial relationships among pixels.The KFDAIMED features of BMP2sn-9563 and the training set are shown in Fig.6.The features of three training sample sets(BMP2sn-c21,BTR70sn-c71 and T72sn-132)are separated welland the features of BMP2sn-9563 are also gathered well around BMP2sn-c21.Fig.6 demonstrates that the features can be extracted by KFDA-IMED robustly.In addition,the kernel parameters for KFDA-IMED can be calculated with the samples,which demonstrates its superioradaptability.

Fig.6 KFDA-IMED features of BMP2sn-9563 and the training set

The vehicles may appear in a range of complicated environments.In order to demonstrate the in fl uence of size of image chip on the recognition results,all serials of vehicles in different sizes are recognized with the KFDAIMED+SVMapproach.

The recognition results are listed in Table 2.

Table 2 Recognition results ofvehicles in different sizes%

Itcan be seen from Table 2 that,for same serialtargets, reducing the size of the image chip can lead to a gradual increase in the recognition rate.For the case of varianttargets,reducing the size of the image chip can make the recognition rate gradually increase at fi rst before starting to fall.The turning pointor threshold appears to be the image chip size 96×96.For a more detailed analysis of the experimental data,we show T72sn-812 in three different sizes(128×128,96×96,32×32)in Fig.7.We can see that the pixels of both the targetand the shadow are completely preserved if the image size is largerthan 96×96.When the image size is smaller than 96×96,the shadow information is lost.When the image size is reduced to 32×32,a smallpart of the target pixels are lost as well.Fig.7 combined with Table 2,shows that for the same serialtargets,when the vehicles themselves have no difference,only relying on the target pixels leads to good recognition results.If most of the target pixels are contained in the image chip, the size change has little effecton the recognition rate but too many background clutter pixels are likely to cause interference.When the target pixels and shadow pixels are all contained in the image chip,reducing the background clutterpixels willimprove the recognition rate forthe varianttargets.When the shadow information is lostgradually, the recognition rate begins to decrease.Our experimental results demonstrate thatimportantdiscriminative information is contained in shadow pixels,which can help recognize variant targets.Further,in order to avoid excessive interference,the irrelevantbackground clutter pixels need to be reduced.

Fig.7 T72sn-812 image chip in three different sizes

4.2 Vehicle detection and recognition experiment

In this experiment,nine vehicle chips(at 15°depression angle)are inserted into a 1 478×1 784 fullclutter MSTAR SAR image[5].Fig.8 shows the image with nine vehicles. White frames are used to indicate the locations of vehicles and serial numbers of the nine vehicles,one of which is seen on the road,two on grass and the other six in groves.

Fig.8 MSTAR SAR image with nine vehicles

Our proposed KFDA algorithm is compared with the classic CFAR algorithm for vehicle detection.In our method,the selection of the shape and size of the sliding window is mainly based on the generalshape and size of vehicles.According to the analysis of Fig.7,the size of inner-window is setat48×48.In orderto make the number of pixels in the outer-window same as the inner-window, we set the size of the protecting window at 55×55 and the size of the outer-window at 73×73.As for the kernel function,we choose the easier-to-controlpolynomialfunction k(x,y)=(xy+1)l,where l=5.The constantκ is 10?4.The detection result is shown in Fig.9.The nine vehicles are all detected(shown in red)and the number of false alarms is four(shown in white).For CFAR,the detection result is shown in Fig.10.Eight vehicles are detected(shown in red)and the number of false alarms is ten(shown in white).Since CFAR mainly considers the contrast feature and ignores geometric information and features of vehicles,it is possible to generate many false alarms in a complicated environment.It is also hard to detect shaded vehicles because of the weak contrast. However,our KFDA algorithm considers the geometric information,features of vehicles as well as the differencebetween features of vehicles and their surrounding environmentand does notneed to face the challenges of selection and extraction of the features.It has a better description and understanding of the vehicles and can produce a higher detection rate and lowerfalse alarm rate.

Fig.9 Detection results of the algorithm based on KFDA(red:vehicles,white:false alarms)

Fig.10 Detection results of CFAR(red:vehicles,white:false alarms)

After detection,the geometric center of each target is locked again to collectthe 48×48 image chips.The image chip of the second vehicle is shown in Fig.11.Itis obvious that a lotof pixels of trees are mixed in the image chip.It may lead to a substantialinterference with the recognition. According to the analysis of Table 2,the size ofimage chip is reduced to 32×32 and our KFDA-IMED approach and trained SVMin experiment4.1 are employed to recognize the targets.The recognition results are shown in Fig.12 for the detected targets.“Red”represents correct recognition results and“black”represents wrong results.Their correctserial numbers are indicated in parentheses.Seven vehicles are correctly recognized and the recognition rate is 77.78%.For the four false alarms,recognition results from leftto rightare:BTR70,BMP2,BMP2 and BTR70. As mentioned about,some environmentsamples in certain areas can be collected,which are around yet do notcover the H1image chips.The four false alarms now are recognized as background,background,BMP2 and background respectively.Meanwhile,the recognition results of the nine vehicles remain unchanged.This process illustrates that the recognition algorithm can help eliminate three false alarms.

Fig.11 48×48 image chip ofthe second vehicle

Fig.12 Recognition results of KFDA-IMED+SVM(Red:correct recognition results;Black:wrong recognition results whose correct serialnumbers are indicated in parentheses)

There are two main reasons why the recognition rate in this experiment is lower than that in experiment 4.1. First,part of the targetpixels and shadow information are lostas reducing the size of image chips,which may result in a lower recognition rate.Second,because most vehicles are in the groves,they are partially shaded by trees. This dramatically changes and masks the pixels of vehicles and their shadows,which increases the dif fi culties of accurate recognition.For the three clearly exposed targets, the recognition results are allseen to be correct.These experimentalresults conform to our theoreticalanalysis.

5.Conclusions

Targetdetection and recognition is a key issue in SAR image applications.By analyzing the mathematicalprinciples of KFDA and relevant features of vehicles,we propose a vehicle detection method based on KFDA and a vehicle recognition method based on the improved KFDA-IMED. Experimental results with the MSTAR database demonstrate.First,our detection method does notneed any prior information and avoids the dif fi culty of selecting features and extracting advanced features,which proves the advantages of introducing KFDA into detection.Second,our method performs better than the traditional CFAR algorithm when vehicles are shaded in a complicated environment.Third,KFDA-IMED+SVM is an ideal SAR image feature extraction and recognition algorithm.Although the vehicles include variations in aspectangle,targetcon fi gurations and obscuration and the clutter interference is excessive,KFDA-IMED+SVMdoes notneed image preprocessing and image segmentation and can lead to a higher recognition rate than conventionalmethods.Fourth,the detection and recognition are achieved by the same criterion. In addition,collecting background clutter samples in the recognition partcan help reduce the false alarms in the detection part.In this paper,a detailed discussion is also presented on the in fl uence of size of image chip on the recognition rate.In general,the simplicity,innovation and the signi fi cant effect of our vehicle detection and recognition algorithm based on KFDA meets the requirements of SAR ATR.Itis concluded to be a highly ef fi cientvehicle detection and recognition algorithm for SAR images.

[1]L.M.Novak,M.C.Burl,W.W.Irving.Optimal polarimetric processing for enhanced target detection.IEEE Trans.on Aerospace and Electronic Systems,1993,29(1):234-244.

[2]D.Wang.Knowledge based object detection using SAR im-ages.Proc.ofthe IEEE InternationalConference on Acoustics, Speech,and SignalProcessing,1987:269-272.

[3]D.Blacknell.Contextual information in SAR target detection.IEE Proceedings of Radar,Sonar and Navigation,2001, 148(1):41-47.

[4]W.W.Irving,L.M.Novak,A.S.Willsky.A multiresolution approach to discrimination in SAR imagery.IEEE Trans.on Aerospace and Electronic Systems,1997,33(4):1157-1169.

[5]P.Lombardo,M.Sciotti,L.M.Kaplan.SAR prescreening using both target and shadow information.Proc.of the IEEE Radar Conference,2001:147-152.

[6]S.Liu,Z.Cao.SAR image target detection in complex environments based on improved visual attention algorithm. EURASIP Journal on Wireless Communications and Networking,2014(1):1-8.

[7]B.Bhanu.Automatic target recognition:state of the art survey.IEEE Trans.on Aerospace and Electronic Systems,1986, AES-22(4):364-379.

[8]B.Bhanu,T.L.Jones.Image understanding research for automatic target recognition.IEEE Aerospace and Electronic Systems Magazine,1993,8(10):15-23.

[9]B.Bhanu,D.E.Dudgeon,E.G.Zelnio,etal.Guesteditorial introduction to the specialissue on automatic targetdetection and recognition.IEEE Trans.on Image Processing,1997,6(1): 1-6.

[10]T.D.Ross,S.Worell,V.Velten,et al.Standard SAR ATR evaluation experiments using the MSTAR public release data set.International Society for Optics and Photonics in Aerospace/Defense Sensing and Controls,1998,3370:566-573.

[11]M.Hurst,R.Mittra.Scattering center analysis via Prony's method.IEEE Trans.on Antennas and Propagation,1987, 35(8):986-988.

[12]L.C.Potter,R.L.Moses.Attributed scattering centers for SAR ATR.IEEE Trans.on Image Processing,1997,6(1):79-91.

[13]T.D.Ross,J.J.Bradley,L.J.Hudson,et al.SAR ATR:so what's the problem?Proc.of the SPIE Conference on Algorithms for Synthetic Aperture Radar Imagery,1999:662-672. [14]Y.Lin,B.Bhanu.Evolutionary feature synthesis for object recognition.IEEE Trans.on Systems,Man,and Cybernetics, Part C:Applications and Reviews,2005,35(2):156-171.

[15]X.Lu,P.Han,R.Wu.Two-dimensional PCA for SAR automatic target recognition.Proc.of the 1st Asian and Pacific Conference in Synthetic Aperture Radar,2007:513-516.

[16]P.Han,K.B.Wu,Z.H.Wang,et al.SAR automatic target recognition based on KPCA criterion.Journal of Electronics and Information Technology,2003,25(10):1297-1301.

[17]H.S.Seung,D.D.Lee.The manifold ways ofperception.Science,2000,290(5500):2268-2269.

[18]M.L.Bryant,F.D.Garber.SVM classi fi er applied to the MSTAR public data set.International Society for Optics and Photonics in Aerosense,1999:355-360.

[19]Z.Qun,J.C.Principe.Supportvectormachines for SAR automatic targetrecognition.IEEE Trans.on Aerospace and Electronic Systems,2001,37(2):643-654.

[20]Y.Sun,Z.Liu,S.Todorovic,et al.Synthetic aperture radar automatic target recognition using adaptive boosting.International Society for Optics and Photonics in Defense and Security,2005:283-293.

[21]D.C.Tao,X.L.Li,X.D.Wu,etal.Geometric mean forsubspace selection.IEEE Trans.on Pattern Analysis and Machine Intelligence,2009,31(2):260-274.

[22]G.Camps-Valls,L.Bruzzone.Kernelmethods for remote sensing data analysis.WestSussex:Wiley Online Library,2009.

[23]V.Vapnik.The nature ofstatisticallearning theory.New York: Springer,2000.

[24]L.W.Wang,Y.Zhang,J.F.Feng.On the Euclidean distance of images.IEEE Trans.on Pattern Analysis and Machine Intelligence,2005,27(8):1334-1339.

[25]J.Brewer.Kronecker products and matrix calculus in system theory.IEEE Trans.on Circuits and Systems,1978,25(9): 772-781.

[26]L.P.Hu.Investigation on targetrecognition from SAR images. Xi'an:Xidian University,2009.(in Chinese)

[27]M.Amoon,G.A.Rezai-rad.Automatic target recognition of synthetic aperture radar(SAR)images based on optimalselection of Zernike momentsfeatures.IET Computer Vision,2014, 8(2):77-85.

[28]U.Srinivas,V.Monga,R.G.Raj.SAR automatic targetrecognition using discriminative graphical models.IEEE Trans.on Aerospace and Electronic Systems,2014,50(1):591-606.

[29]P.L.Douville.Measured and predicted synthetic aperture radar target comparison.IEEE Trans.on Aerospace and Electronic Systems,2002,38(1):25-37.

Biographies

Fei Gaowas born in 1975.He received his Ph.D. degree in signals and information processing from Beihang University in 2005.He is currently an associate professor in School of Electronic and Information Engineering,Beihang University.His research interests include radarsignalprocessing,image processing,and moving targetdetection.

E-mail:feigao2000@163.com

Jingyuan Meiwas born in 1990.He received his B.E.degree in Information Engineering Schoolfrom Communication University of China.He is currently a master in School of Electronics and Information Engineering,Beihang University.His research interests include radar signalprocessing,image processing,target detection,target recognition,and online learning.

E-mail:347068642@qq.com

Jinping Sunwas born in 1975.He received his Ph.D.degree in signals and information processing from Beihang University in 2001.He is currently a professor of School of Electronic and Information Engineering,Beihang University.His research interests include high resolution and advance mode SAR signalprocessing,image understanding and pattern recognize.

E-mail:sunjinping@buaa.edu.cn

Jun Wangwas born in 1972.He received his Ph.D.degree in signals and information processing from Beihang University in 2001.He is currently a professor in School of EIE,Beihang University.His research interests include signalprocessing,DSP/FPGAreal-time architecture,targetrecognition and tracking.

E-mail:wangj203@buaa.edu.cn

ErfuYangreceived his B.E.,M.E.,and Ph.D.degrees from Beihang University in 1994,1996,and 1999,respectively.He is currently a lecturer in University of Strathclyde.His research interests include signalprocessing and cognitive signal-image.

E-mail:erfu.yang@strath.ac.uk

AmirHussainobtained his B.Eng.(with 1st Class Honors)and Ph.D.degrees,both in electronic and electrical engineering from the University of Strathclyde in Glasgow,Scotland,in 1992 and 1996,respectively.He is currently a professor in School of Natural Sciences,University of Stirling and founding director ofthe Cognitive Signal-image and Control Processing Research(COSIPRA)Laboratory. His research interests are mainly interdisciplinary,and include machine learning and cognitive computing for modeling and control of complex systems.

E-mail:kfa@cs.stir.ac.uk

10.1109/JSEE.2015.00080

Manuscriptreceived September 18,2014.

*Corresponding author.

This work was supported by the National Natural Science Foundation of China(61071139;61471019;61171122),the Aeronautical Science Foundation of China(20142051022),the Foundation of ATR Key Lab(C80264),the National Natural Science Foundation of China(NNSFC)under the RSE-NNSFC Joint Project(2012-2014) (61211130210)with Beihang University,the RSE-NNSFC Joint Project (2012-2014)(61211130309)with Anhui University,and the“Sino-UK Higher Education Research Partnership for PhD Studies”Joint Project (2013-2015).