CN101135729A - Method for synthetic aperture radar shelter from shelter from object identification based on supporting vector machine - Google Patents
Method for synthetic aperture radar shelter from shelter from object identification based on supporting vector machine Download PDFInfo
- Publication number
- CN101135729A CN101135729A CNA2007100185741A CN200710018574A CN101135729A CN 101135729 A CN101135729 A CN 101135729A CN A2007100185741 A CNA2007100185741 A CN A2007100185741A CN 200710018574 A CN200710018574 A CN 200710018574A CN 101135729 A CN101135729 A CN 101135729A
- Authority
- CN
- China
- Prior art keywords
- target
- image
- shelter
- roi
- sorter
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Landscapes
- Radar Systems Or Details Thereof (AREA)
Abstract
The method is used for overcome the falling of the recognition rate when the shielded rate of the target is over 70%. It comprises: 1) making pretreatment for the SAR target image to be recognized to get the ROI of the processed image; searching the peak point of the target in the ROI of the processed image; for the processed ROI, removing a part of peak point of the target; using type I or type II approach to add a corresponding numbers of peak points to the ROI in order to shield the target; selecting the classifier, and training the classifier; adding the test image to the trained classifier.
Description
Technical field
The invention belongs to technical field of image processing, relate to this technology at synthetic-aperture radar Synthetic Aperture Radar, it is the application in SAR image object identification field, specifically a kind of based on supporting vector machine Support VectorMachine, the SAR shelter target recognition methods of SVM.This method can be used for important goals such as tank, cannon, pillbox by in the identification of the military target under the partial occlusion situation.
Background technology
The ability that SAR has is round-the-clock, round-the-clock is observed earth surface, and raising along with signal processing technology, reached at present very high spatial resolution, this makes the SAR sensor bring into play increasing effect in military affairs investigation and battlefield perception, and SAR automatic target identification at present is a field that is subjected to showing great attention to both at home and abroad.In the identification of military target, the targets such as tank, panzer, cannon, bunker that run into the enemy through regular meeting are through camouflage and situation about being blocked, and the recognition methods of therefore studying shelter target in the SAR image is highly significant.
Because the scene of SAR observation is generally bigger, so the general method that adopts layering of SAR image object identification: at first extract area-of-interest Regions of Interest from whole SAR image, ROI promptly comprises target all or the zone of major part; Then these interesting areas are done further processing, reaching the classification to target, the purpose of identification, thereby improve the efficient of target identification system greatly.
Traditional SAR target identification method mostly adopts global characteristics that image is discerned.These class methods are at first extracted eigenvector from the ROI image, and comparing with the sample that target database prestores by the eigenvector that extracts then reaches the purpose of target classification.And in the target of blocking was arranged, owing to block the existence of situation, bigger variation can take place as shape, profile, main shaft etc. in the global characteristics of target, so these methods are for shelter target and inapplicable.
People such as Bhanu proposed to extract the peak value feature of SAR image in 1999, simulate the SAR shelter target by peak value characteristic sequence structure question blank, by local feature point coupling the clauses and subclauses in the question blank are voted then, poll surpasses the ballot threshold value and just determines its affiliated classification, and this method abbreviates the Bhanu method in the back as.The Bhanu method is to blocking, and the SAR Target Recognition effect that the rate of especially blocking is not high is fine, and can reduce to suddenly about 50% up to 70%～90% o'clock discrimination in the rate of blocking, and almost can not work.
Owing to also there is not the public database of SAR shelter target at present, therefore the research to the identification of SAR shelter target need simulate SAR shelter target data earlier.People such as Bhanu proposed relatively reasonably SAR shelter target analogy method in 1999, yet the data that this method simulates have cooperated the needs of the recognizer of their propositions, compared with the SAR shelter target of reality also to have certain irrationality.
Summary of the invention
The objective of the invention is to: in order to overcome the deficiency of prior art, promptly discrimination can descend rapidly when the rate of blocking surpasses 70%, a kind of SAR shelter target recognition methods based on supporting vector machine has been proposed, it had both guaranteed the low high discrimination that blocks under the rate, make the discrimination that blocks under the rate at height have greatly improved again, have robustness preferably.
Technical scheme of the present invention is: supporting vector machine is applied in the identification of SAR shelter target as sorter.At first to the motion, static target is obtained and discern Moving and Stationary Target Acquisition and Recognition, the MSTAR data are handled, simulate shelter target, therefrom take out 17 ° of depression angle images as training image, be divided into one group according to its azimuth information since 0 ° of per 10 ° of interval, being divided into is 36 groups, accordingly 36 sorters is trained; Then the image that takes out 15 ° of depression angles from the MSTAR data is discerned with corresponding azimuthal sorter that trains according to its azimuth information as test pattern.Technical scheme specific implementation process of the present invention is as follows:
(1), the SAR target image to be identified of input is carried out pre-service, obtain the ROI of image after the pre-service;
(2), after pre-service, seek the peak point of target among the ROI of image;
(3), to pretreated ROI, by removing the part peak point of target, the peak point that adds respective number with I class or II class methods is simulated shelter target in ROI;
(4), select sorter for use and sorter trained, in the sorter that the test pattern input is trained, carry out shelter target identification.
Above-mentioned a kind of method for synthetic aperture radar shelter from shelter from object identification based on supporting vector machine, said SAR target image to be identified to input carries out pre-service, obtains the ROI of image after the pre-service, and its method is as follows:
(1), with median filter to the input picture denoising, to weaken the influence of speckle noise;
(2), image is carried out normalization, image pixel value is normalized to 0～1;
(3), image after the normalization is carried out Threshold Segmentation, the threshold value of cutting apart is the standard deviation that the average of image adds twice after the normalization, is generally 0.8;
(4), image after the Threshold Segmentation is carried out the mathematical morphology conversion: the method that employing is expanded makes the interregional slit that splits be filled, and corrodes to make the zone become the target main areas of image and several zonules on every side again.
(5), the wavenumber filter of using tricks carries out filtering, setting counting wave filter size is 5 * 5, is in the image-region of 5 * 5 sizes at center with certain pixel promptly, if pixel value be zero number of pixels greater than 5 * 5/2, then this pixel value is adjusted into zero, otherwise, do not adjust;
(6), carry out expansive working, remedy some bright spots that the target area edge penalty falls, promptly at the non-vanishing point of some pixel values of target area edge supplement.
(7), find out the position of the non-vanishing point of the pixel value of (6) resultant image, to in the input picture pixel that should the position be kept initial value, the pixel value zero setting of rest of pixels, taking out size from the input picture center position then is 54 * 54 subgraph, has just obtained the ROI of input picture.
Above-mentioned a kind of method for synthetic aperture radar shelter from shelter from object identification based on supporting vector machine, the said peak point of after pre-service, seeking target among the ROI of image, its method is to search for successively according to raster scan order, if this point is to be the point that has max pixel value in 3 * 3 image-regions at center with it, then this is a peak point, writes down its pixel value and coordinate; Otherwise continue search.
Above-mentioned a kind of method for synthetic aperture radar shelter from shelter from object identification based on supporting vector machine, said to pretreated ROI, by removing the part peak point of target, add the peak point of respective number in ROI with I class or II class methods, simulate shelter target, its process is as follows:
(1), determine after the target peak point of ROI, from four direction: against the orientation of aircraft to d
1, go up distance to d
2, along the orientation of aircraft to d
3With following distance to d
4In either direction begin, remove and run into successively, number is the peak point of peak point sum N%, simulates the shelter target that the rate of blocking is N%, N ∈ [0,100].
(2), the random site that the pixel of number same as described above, random value is added the zone that is deleted peak point.In the present invention, adopt I class or two kinds of methods of II class to simulate: the pixel that adds in the I class methods is obeyed the even distribution in [0,1]; The pixel that adds in the II class methods is obeyed the even distribution in [0.8,1];
Above-mentioned a kind of method for synthetic aperture radar shelter from shelter from object identification based on supporting vector machine is saidly selected sorter for use and sorter is trained, and in the sorter that the test pattern input is trained, carries out shelter target identification.Wherein saidly select sorter for use and sorter is trained, its implementation procedure is: on a surface target imaging data was as training in the training sample input supporting vector machine when MSTAR was 17 ° in the depression angle, whole training datas are divided into one group according to its azimuth information since 0 ° of per 10 ° of interval, being divided into is 36 groups, accordingly 36 sorters is trained; According to the training data difference, adopt following two kinds of methods to train:
1., adopt single training set, promptly include only unscreened image to be identified in the training set;
2., adopt many training sets, promptly both comprised unscreened image to be identified in the training set, comprise again with image to be identified and simulate the shelter target image of the rate of blocking under 10%～90% situation;
In the said sorter that test pattern input is trained, discern, its implementation procedure is: on a surface target imaging data took out as test pattern when MSTAR was 15 ° in the depression angle, and, just can finish identification to shelter target with in the above-mentioned counterparty's parallactic angle sorter that has trained of this test pattern input.
The present invention has the following advantages compared with prior art:
1 the present invention is very effective to the especially high rate Target Recognition of blocking of SAR shelter target, has robustness preferably;
The SAR image preprocessing process that 2 the present invention adopt has guaranteed the degree of accuracy of follow-up identification;
3 the present invention is directed to the relation of different target and background, adopt two kinds of simulation shelter target methods, have guaranteed that the target information disappearance is less.
Description of drawings
Fig. 1 is a realization flow synoptic diagram of the present invention
Fig. 2 is that the present invention is to each step results synoptic diagram of SAR image pre-service
Fig. 3 is the different target image synoptic diagram that block under the rate of input picture master drawing, ROI and simulation among the present invention
Fig. 4 is that the present invention adopts I class methods OB type operation simulation shelter target, to the average recognition result of whole test patterns and the recognition result synoptic diagram of different classes of test pattern
Fig. 5 is that the present invention adopts II class methods OB type operation simulation shelter target, to the average recognition result of whole test patterns and the recognition result synoptic diagram of different classes of test pattern
Fig. 6 is that the present invention adopts OB type operation simulation shelter target, and I class methods and II class methods are to the total discrimination of whole test patterns and the comparison synoptic diagram of Bhanu method recognition result
Fig. 7 is that the present invention is with I class methods BB type operation simulation shelter target, to the total recognition result of whole test patterns and the recognition result synoptic diagram of different classes of test pattern
Fig. 8 is that the present invention is with II class methods BB type operation simulation shelter target, to the total recognition result of whole test patterns and the recognition result synoptic diagram of different classes of test pattern
Fig. 9 is the comparison synoptic diagram of BB type and OB type operation simulation shelter target and Bhanu method recognition result
Embodiment
With reference to Fig. 1, it is a realization flow synoptic diagram of the present invention, and as can be seen from the figure its specific implementation process is as follows:
1, the SAR target image to be identified to input carries out pre-service, and the pre-service that obtains the ROI image of image after the pre-service can be finished with following several steps:
(1), with median filter to the input picture denoising, to weaken the influence of speckle noise;
(2), image after the denoising is carried out normalization, image pixel value is normalized to 0～1;
(3), image after the normalization is carried out Threshold Segmentation, the threshold value of cutting apart is the standard deviation that the average of image adds twice after the normalization, is generally 0.8;
(4), image after the Threshold Segmentation is carried out the mathematical morphology conversion: the method that employing is expanded makes the interregional slit that splits be filled, and corrodes to make the zone become the target main areas of image and several zonules on every side again.Wherein should note expanding and the control of extent of corrosion, be as the criterion as far as possible just to cover former target;
(5), the wavenumber filter of using tricks carries out filtering, setting counting wave filter size is 5 * 5, is in the image-region of 5 * 5 sizes at center with certain pixel promptly, if pixel value be zero number of pixels greater than 5 * 5/2, then this pixel value is adjusted into zero, otherwise, do not adjust;
(6), carry out expansive working, remedy some bright spots that the target area edge penalty falls, promptly at the non-vanishing point of some pixel values of target area edge supplement.When carrying out expansive working, should note the control of degrees of expansion, be as the criterion as far as possible just to cover former target;
(7), find out the position of the non-vanishing point of the pixel value of (6) resultant image, to in the input picture pixel that should the position be kept initial value, the pixel value zero setting of rest of pixels, taking out size from the input picture center position then is 54 * 54 subgraph, has just obtained the ROI of input picture.The SAR image that is used to handle among the present invention all is through calibrating, target all is positioned at the center of image, here take out size from the picture centre position and be 54 * 54 the subgraph ROI as input picture, purpose is that all the other background areas are removed, thereby reduces background interference and reduce operand.
Fig. 2 is each intermediate steps result schematic diagram of SAR image preprocessing process among the present invention, Fig. 2 (a) is the SAR target image to be identified of input, Fig. 2 (b) is to the figure as a result after the input picture denoising, Fig. 2 (c) is the figure as a result that image after the normalization is carried out Threshold Segmentation, Fig. 2 (d) is expand figure as a result after the corrosion of the figure as a result to Fig. 2 (c), Fig. 2 (e) is to the filtered figure as a result of Fig. 2 (d) counting, and Fig. 2 (f) is the ROI figure as a result of the input picture that obtains after pre-service is finished to input picture.
2, after pre-service, seek the peak point of target among the ROI of image
According to the point of raster scan order search peak successively, if this point is to be the point that has max pixel value in 3 * 3 image-regions at center with it, then this is a peak point, writes down its pixel value and coordinate; Otherwise continue search.
The step of extracting peak point among the present invention is served for following simulation shelter target.After it should be noted that the present invention has determined the peak point of image, do not require that different images unifies the peak point number.This is because azimuthal difference, same model target can have the peak point of different numbers, if unified is bigger number, then the less target of number will be replenished " pseudo-peak value " originally, these " pseudo-peak values " are non-existent in fact, and peak value real after the rate of blocking acquires a certain degree just may all have been blocked, and in fact all blocked such target this moment, increase the rate of blocking again to them and Yan Buhui has any different, such situation is also unreasonable; If unified be less number, then the more target of number will lose some peak points originally, and the more target of some peak point may lose most of peak value, so just can't describe target exactly, also is unfavorable for improving discrimination.
3, to pretreated ROI, by removing the part peak point of target, the peak point of using I class or II class methods interpolation respective number is simulated shelter target in ROI
The process of simulation shelter target is as follows among the present invention:
(1), determine after the target peak point of ROI, from four direction: against the orientation of aircraft to d
1, go up distance to d
2, along the orientation of aircraft to d
3With following distance to d
4In either direction begin, remove and run into successively, number is the peak point of peak point sum N%, simulates the shelter target that the rate of blocking is N%, N ∈ [0,100].It should be noted that, direction selected when removing peak point has generality, because all there is distribution at the target direction angle of each model in the MSTAR data in 0 °～360 ° scopes, therefore select the different directions in the four direction only can produce minor impact to recognition result in theory, the present invention chooses down distance to direction;
(2), the random site that the pixel of number same as described above, random value is added the zone that is deleted peak point.The random value scope of the pixel of these addings can exert an influence to recognition result.Because the segmentation threshold in the step 1 is made as 0.8, if the some pixel value that adds is less than 0.8, so these " secretly " name a person for a particular job and can not be counted as the interior point of target zone, also can not be considered to describe the peak point of object construction feature; If the some pixel value that adds is greater than 0.8, these " bright " are named a person for a particular job and are counted as impact point so, and may be counted as the peak point of target.In the present invention, adopt I class or two kinds of methods of II class to simulate: the pixel that adds in the I class methods obeys [0,1] Nei even distribution, it is the pixel value completely random, the number of point generally can reduce in the ROI of target in this case, the information of target can lack more, and the interference that target is subjected to is bigger; The pixel that adds in the II class methods is obeyed the even distribution in [0.8,1], and promptly the point of Jia Ruing is " bright " point, and guarantees that by check the number that target is put remains unchanged in ROI, and the loss of learning of target is less, and the interference that target is subjected to is also smaller.
4, select sorter and sorter trained for use, in the sorter that the test pattern input is trained, carry out shelter target identification
On a surface target imaging data is as training in the training sample input supporting vector machine when among the present invention MSTAR being 17 ° in the depression angle, the MSTAR data comprise three major types: BMP2 armo(u)red carrier, BTR70 armo(u)red carrier, T72 main battle tank, various model target direction angular coverages are 0 °～360 °, in order effectively to improve discrimination, whole training datas are divided into one group according to its azimuth information since 0 ° of per 10 ° of interval, being divided into is 36 groups, accordingly 36 sorters is trained.1., adopt single training set according to the training data difference, adopt two kinds of methods to train:, promptly to include only unscreened image to be identified in the training set; 2., adopt many training sets, promptly both comprised unscreened image to be identified in the training set, comprise again with image to be identified and simulate the shelter target image of the rate of blocking under 10%～90% situation; Because the depression angle difference of image, the shelter target when therefore adopting many training sets in the training sample also is different from shelter target in the test sample book.
On a surface target imaging data was used for doing test data when MSTAR was 15 ° in the depression angle.For test pattern, discern with corresponding azimuthal sorter according to its azimuth information.In the above-mentioned counterparty's parallactic angle sorter that has trained of this test pattern input, just can finish identification to shelter target.
The present invention selects k nearest neighbor K-Nearest Neighbor for use, and two kinds of sorters of KNN and SVM are discerned.Select parameter when it is pointed out that 36 groups of training trained sorters respectively.Parameter selection aspect: the value of K is [1,3,5,7,9] among the KNN, and what SVM used is linear support vector machine, and C gets 2
17～2
21, adopt 10 times of cross validation methods to select the C value.
Fig. 3 (a) is a SAR target master drawing, Fig. 3 (b) is the ROI of Fig. 3 (a), the present invention is adopting two kinds of methods of operating in the l-G simulation test: a kind of shown in Fig. 3 (c), on the basis of Fig. 3 (a), simulate, the shade of target and background parts all keep initial value, here called after OB type Original Background; Another kind of shown in Fig. 3 (d), the simulation of on the basis of Fig. 3 (b), blocking, the simulated operation in the target area is identical with first kind, just with background and all zero setting of target shadow zone, promptly only keep the point in the target area, here called after BB type Black Background.Fig. 3 (c) is an OB type method of operating, and last row is for adopting I class methods simulation 30%, 60%, the 90% shelter target image that blocks under the rate, and following row is for adopting above-mentioned three shelter target analog images that block under the rate of II class methods simulation.Fig. 3 (d) is a BB type method of operating, is simulation 30%, 60%, the 90% shelter target image that blocks under the rate equally, and last row is for adopting the I class methods, and following row is for adopting the II class methods.
Table 1 has provided with 15 ° of depression angle images in the MSTAR data and has made test data, adopt I class methods OB type operation simulation shelter target, the experimental result of using SVM and KNN sorter that tertiary target BMP2 armo(u)red carrier, BTR70 armo(u)red carrier and T72 main battle tank are discerned is comprising every classification target discrimination and average recognition rate.Fig. 4 is and the corresponding discrimination curve of table 1, Fig. 4 (a) is that the test data rate of blocking was simulated 90% o'clock from 0%, adopt the KNN and the svm classifier device of single training set training, adopt the KNN of many training set training and the average recognition rate curve map that the svm classifier device is discerned BMP2 armo(u)red carrier, BTR70 armo(u)red carrier and T72 main battle tank tertiary target
Table 1 I class methods average recognition rate and each identification of targets rate (OB type)
| Sorter | Test data is blocked rate | (1) single training set | (2) many training sets | ||||||
| BMP2 | BTR70 | T72 | Average recognition rate | BMP2 | BTR70 | T72 | Average recognition rate | ||
| KNN | 0 10％ 20％ 30％ 40％ 50％ 60％ 70％ 80％ 90％ | 100％ 99.8％ 99.8％ 99.3％ 99.2％ 98.8％ 98.0％ 96.6％ 94.9％ 92.7％ | 98.0％ 97.5％ 92.9％ 83.7％ 73.5％ 69.9％ 66.8％ 58.7％ 47.5％ 40.8％ | 96.1％ 95.4％ 91.6％ 86.4％ 80.2％ 73.4％ 67.2％ 63.1％ 57.0％ 52.9％ | 98.02％ 97.58％ 95.31％ 91.58％ 87.40％ 83.81％ 80.37％ 76.85％ 71.94％ 68.28％ | 100％ 99.8％ 99.8％ 99.7％ 99.2％ 99.0％ 98.3％ 97.3％ 96.1％ 92.7％ | 98.5％ 96.9％ 94.4％ 87.2％ 79.6％ 71.4％ 69.9％ 63.3％ 55.6％ 48.5％ | 96.1％ 95.4％ 91.4％ 84.5％ 76.5％ 72.0％ 62.5％ 60.0％ 54.8％ 49.7％ | 98.10％ 97.51％ 95.46％ 91.43％ 86.67％ 83.52％ 78.97％ 76.48％ 72.67％ 67.99％ |
| SVM | 0 10％ 20％ 30％ 40％ 50％ 60％ 70％ 80％ 90％ | 99.8％ 99.8％ 99.8％ 99.7％ 99.3％ 99.2％ 98.5％ 98.5％ 96.3％ 95.9％ | 96.4％ 91.8％ 82.1％ 69.4％ 61.2％ 56.6％ 59.7％ 58.2％ 48.0％ 40.3％ | 97.3％ 95.4％ 92.6％ 85.4％ 78.0％ 70.6％ 64.8％ 58.1％ 52.2％ 45.2％ | 98.24％ 96.78％ 94.21％ 89.23％ 84.76％ 80.88％ 78.53％ 75.46％ 70.55％ 66.30％ | 99.5％ 99.5％ 99.5％ 99.3％ 99.5％ 99.5％ 99.5％ 99.5％ 99.5％ 99.5％ | 87.8％ 84.7％ 82.7％ 82.1％ 81.6％ 80.6％ 76.0％ 75.0％ 68.4％ 64.8％ | 96.1％ 96.1％ 96.6％ 96.2％ 96.2％ 95.5％ 95.2％ 94.3％ 90.6％ 92.4％ | 96.34％ 95.90％ 95.82％ 95.53％ 95.53％ 95.09％ 94.29％ 93.77％ 92.60％ 91.50％ |
Hereinafter use these four kinds the sorter of distinct methods training to abbreviate four kinds of contrast sorters as, the horizontal ordinate among the figure is the rate of blocking of test data, and ordinate is illustrated in this and blocks under the rate, the correct recognition rata that uses sorter to discern.The discrimination curve map of Fig. 4 (b)～(d) for using above-mentioned four kinds of contrast sorters respectively three kinds of targets to be discerned.From Fig. 4 (a)～(b) all as can be seen, in the rate of blocking greater than 50% o'clock, the SAR shelter target recognition methods that the present invention proposes based on supporting vector machine, when adopting many training sets method training svm classifier device, the discrimination curve is all apparently higher than other method, and curve is steady, demonstrates method proposed by the invention and has high discrimination under the rate low blocking, the discrimination that blocks under the rate at height also has greatly improved than other method, has robustness preferably.As can be seen from Table 1, the rate of blocking of test data was simulated 90% o'clock from 0%, when adopting many training sets method training svm classifier device, the svm classifier device for the average recognition rate of BMP2 armo(u)red carrier, BTR70 armo(u)red carrier, T72 main battle tank all more than 90%, significantly better than the KNN sorter.
The confusion matrix of table 2 I class methods (OB type). (a) 20% result of blocking under the rate, (b) 70% result of blocking under the rate
| 20% test data of blocking | (1) single training set | (2) many training sets | |||||
| BMP2 | BTR70 | T72 | BMP2 | BTR70 | T72 | ||
| KNN | BMP2 BTR70 T72 | 586 11 49 | 0 182 0 | 1 3 533 | 586 9 50 | 0 185 0 | 1 2 532 |
| SVM | BMP2 BTR70 T72 | 586 29 43 | 0 161 0 | 1 6 539 | 584 22 20 | 0 162 0 | 3 12 562 |
(a)
| 70% test data of blocking | (1) single training set | (2) many training sets | |||||
| BMP2 | BTR70 | T72 | BMP2 | BTR70 | T72 | ||
| KNN | BMP2 BTR70 T72 | 567 67 193 | 12 115 22 | 8 14 367 | 571 67 216 | 10 124 17 | 6 5 349 |
| SVM | BMP2 BTR70 T72 | 578 78 230 | 3 114 14 | 6 4 338 | 584 33 33 | 0 147 0 | 3 16 549 |
(b)
Table 2 provided test data the rate of blocking be 20% and the rate of blocking be 70% o'clock, use I class methods OB type operation simulation shelter target, to tertiary target BMP2 armo(u)red carrier, BTR70 armo(u)red carrier and T72 main battle tank identified obfuscated matrix.From the confusion matrix of table 2 as can be seen, BMP2 is known into the other side with T72 two class targets by mistake and has been accounted for the overwhelming majority, is 20% o'clock as test data in the rate of blocking, when using the svm classifier device of many training set training, it is BMP2 that 20 test datas among the T72 are known by mistake, it is T72 that 3 test datas among the BMP2 are known by mistake, and BMP2 and T72 are BTR70 by the mistake knowledge, it is BMP2 that 9 test datas among the BTR70 are known by mistake, 2 test datas among the BTR70 are known by mistake and are T72, and therefore this two classifications target peak point of BMP2 and T72 is distributed with to a certain extent similar as can be seen; And the test data that BTR70 is known is by mistake known by mistake more and is BMP2, therefore we can say than T72, BTR70 target and BMP2 more as.
Table 3 has provided with 15 ° of depression angle images in the MSTAR data and has done test data, adopt II class methods OB type operation simulation shelter target, every classification target discrimination and the average recognition rate of using SVM and KNN sorter that tertiary target BMP2 armo(u)red carrier, BTR70 armo(u)red carrier and T72 main battle tank are discerned.Fig. 5 is and the corresponding discrimination curve of table 3, Fig. 5 (a) is that the test data rate of blocking was simulated 90% o'clock from 0%, adopt four kinds of average recognition rate curve maps that the contrast sorter is discerned tertiary target, Fig. 5 (b)～(d) is for using four kinds of discrimination curve maps that the contrast sorter is discerned tertiary target respectively.
The average recognition rate of table 3 II class methods and each identification of targets rate (OB type)
| Sorter | Test data is blocked rate | (1) single training set | (2) many training sets | ||||||
| BMP2 | BTR70 | T72 | Average recognition rate | BMP2 | BTR70 | T72 | Average recognition rate | ||
| KNN | 0 10％ 20％ 30％ 40％ 50％ 60％ 70％ 80％ 90％ | 100％ 99.8％ 99.7％ 99.8％ 99.5％ 99.3％ 99.0％ 98.8％ 98.1％ 97.4％ | 98.0％ 97.5％ 96.9％ 92.9％ 89.3％ 87.2％ 79.6％ 75.5％ 72.5％ 66.3％ | 96.1％ 96.4％ 95.5％ 92.4％ 89.7％ 84.4％ 81.6％ 78.0％ 73.4％ 68.0％ | 98.02％ 98.02％ 97.51％ 95.68％ 93.85％ 91.21％ 88.79％ 86.59％ 83.88％ 80.44％ | 100％ 99.8％ 99.7％ 99.7％ 99.5％ 99.5％ 99.0％ 99.3％ 98.6％ 98.0％ | 98.5％ 98.0％ 95.4％ 93.4％ 91.8％ 88.8％ 80.6％ 79.6％ 80.1％ 68.9％ | 95.5％ 94.9％ 94.7％ 92.1％ 87.5％ 82.5％ 80.6％ 76.1％ 71.0％ 64.8％ | 97.88％ 97.44％ 96.92％ 95.53％ 93.26％ 90.70％ 88.50％ 86.59％ 84.18％ 79.63％ |
| SVM | 0 10％ 20％ 30％ 40％ 50％ 60％ 70％ 80％ 90％ | 99.8％ 99.8％ 99.8％ 99.8％ 99.5％ 99.5％ 99.3％ 99.3％ 99.2％ 99.2％ | 96.4％ 95.4％ 90.8％ 84.7％ 80.1％ 73.5％ 70.4％ 65.8％ 65.8％ 54.6％ | 97.3％ 96.1％ 95.5％ 92.8％ 89.9％ 85.4％ 80.6％ 77.8％ 72.9％ 65.5％ | 98.24％ 97.58％ 96.70％ 94.65％ 92.60％ 89.74％ 87.18％ 85.35％ 83.15％ 78.39％ | 99.5％ 99.5％ 99.5％ 99.5％ 99.49％ 99.49％ 99.49％ 99.49％ 99.49％ 99.32％ | 90.8％ 89.3％ 85.7％ 85.2％ 82.6％ 81.6％ 76.5％ 76.5％ 73.0％ 65.8％ | 96.4％ 96.2％ 96.1％ 96.2％ 95.4％ 94.5％ 94.0％ 93.0％ 92.1％ 90.0％ | 96.92％ 96.63％ 96.04％ 96.04％ 95.31％ 94.80％ 93.85％ 93.41％ 92.53％ 90.55％ |
Table 4 provided test data the rate of blocking be 20% and the rate of blocking be 70% o'clock, use II class methods OB type operation simulation shelter target, to tertiary target BMP2 armo(u)red carrier, BTR70 armo(u)red carrier and T72 main battle tank identified obfuscated matrix.
From Fig. 4 (a) and Fig. 5 (a) as can be seen, the recognition methods that proposes of the present invention is very effective to the especially high rate identification of targets of blocking of SAR shelter target.In I class methods and the II class methods, many training set training SVM average recognition rate when the rate of blocking reaches 90% has all reached more than 80%, exceeds at least ten percentage points of other methods.
When Fig. 6 (a) has provided I class methods OB type operation simulation shelter target, adopt four kinds of contrast sorters and set the Bhanu method of 30 scattering centers and 50 scattering centers, the average recognition rate curve map that tertiary target BMP2 armo(u)red carrier, BTR70 armo(u)red carrier and T72 main battle tank are discerned.When Fig. 6 (b) has provided II class methods OB type operation simulation shelter target, adopt four kinds of contrast sorters and set the Bhanu method of 30 scattering centers and 50 scattering centers, the average recognition rate curve map that tertiary target BMP2 armo(u)red carrier, BTR70 armo(u)red carrier and T72 main battle tank are discerned.
The confusion matrix of table 4 II class methods (OB type). (a) 20% result of blocking under the rate, (b) 70% result of blocking under the rate
| 20% test data of blocking | (1) single training set | (2) many training sets | |||||
| BMP2 | BTR70 | T72 | BMP2 | BTR70 | T72 | ||
| KNN | BMP2 BTR70 T72 | 585 5 26 | 0 190 0 | 2 1 556 | 585 6 30 | 0 187 1 | 2 3 551 |
| SVM | BMP2 BTR70 T72 | 586 15 26 | 0 178 0 | 1 3 556 | 584 18 23 | 0 168 0 | 3 10 559 |
(a)
| 70% test data of blocking | (1) single training set | (2) many training sets | |||||
| BMP2 | BTR70 | T72 | BMP2 | BTR70 | T72 | ||
| KNN | BMP2 BTR70 T72 | 580 44 119 | 1 148 9 | 6 4 454 | 583 37 129 | 1 156 10 | 3 3 443 |
| SVM | BMP2 BTR70 T72 | 583 65 126 | 0 129 3 | 4 2 453 | 584 35 41 | 0 150 0 | 3 11 541 |
(b)
From Fig. 6 (a) and Fig. 6 (b) as can be seen, the rate of blocking of test data is higher than after 70%, set the Bhanu method of 30 scattering centers and 50 scattering centers, the discrimination curve for the treatment of recognition objective descends suddenly, the rate of blocking is that 90% o'clock discrimination drops to about 50%, be increased to 90% the change procedure from 0% in the rate of blocking, the discrimination curve of the svm classifier device that the many training sets that use the present invention to propose are trained is very steady always, its correct recognition rata remains on more than 90%, significantly better than the Bhanu method.Because in the Bhanu method, the separability information of target has only been got the peak value feature of target.Though the presentation of results peak point of Bhanu method has comprised a large amount of separability information of target, can not illustrate that rest of pixels point does not just comprise separability information.When the rate of blocking was not high, correct information was enough to guarantee very high discrimination; And when the rate of blocking was very high, correct separability information seldom was not enough to reach the ballot threshold value, so discrimination decline is very fast, almost can not work.The method that the present invention proposes is then different, and the feature of being got has comprised all separability information of target.Low when blocking rate, owing to also comprise some by the error message of blocking generation in the proper vector, so discrimination is not as the 3hanu method; And when height blocks rate, few although correct peak value feature remains, also have a large amount of other correct separability information existence, so discrimination is still very high.
Table 5 has provided with 15 ° of depression angle images in the MSTAR data and has made test data, adopt I class methods BB type operation simulation shelter target, the experimental result of using SVM and KNN sorter that tertiary target BMP2 armo(u)red carrier, BTR70 armo(u)red carrier and T72 main battle tank are discerned is comprising every classification target discrimination and average recognition rate.Fig. 7 is and the corresponding discrimination curve of table 5, Fig. 7 (a) is that the test data rate of blocking was simulated 90% o'clock from 0%, adopt four kinds of average recognition rate curve maps that the contrast sorter is discerned tertiary target, Fig. 7 (b)～(d) is for using four kinds of discrimination curve maps that the contrast sorter is discerned tertiary target respectively.
The average recognition rate of table 5 I class methods and each identification of targets rate (BB type)
| Sorter | Test data is blocked rate | (1) single training set | (2) many training sets | ||||||
| BMP2 | BTR70 | T72 | Average recognition rate | BMP2 | BTR70 | T72 | Average recognition rate | ||
| KNN | 0 10％ 20％ 30％ 40％ 50％ 60％ 70％ 80％ 90％ | 99.66％ 99.49％ 99.32％ 99.49％ 99.66％ 99.32％ 98.98％ 98.47％ 98.64％ 98.64％ | 96.43％ 95.92％ 95.41％ 94.90％ 93.88％ 95.41％ 93.88％ 95.41％ 94.90％ 93.88％ | 84.19％ 84.71％ 83.68％ 82.65％ 82.82％ 82.13％ 81.44％ 80.58％ 78.18％ 76.80％ | 92.60％ 92.67％ 92.01％ 91.65％ 91.65％ 91.43％ 90.77％ 90.40％ 89.38％ 88.64％ | 99.66％ 99.49％ 99.49％ 99.49％ 99.32％ 99.32％ 98.98％ 98.81％ 98.47％ 98.30％ | 96.43％ 96.94％ 96.43％ 96.43％ 96.43％ 95.92％ 95.41％ 95.92％ 94.90％ 95.41％ | 84.19％ 85.40％ 84.36％ 84.36％ 82.99％ 83.51％ 81.79％ 79.73％ 79.90％ 75.26％ | 92.60％ 93.11％ 92.60％ 92.60％ 91.94％ 92.09％ 91.14％ 90.26％ 90.04％ 88.06％ |
| SVM | 0 10％ 20％ 30％ 40％ 50％ 60％ 70％ 80％ 90％ | 99.32％ 99.32％ 99.49％ 98.98％ 98.81％ 98.64％ 97.96％ 97.96％ 97.96％ 97.96％ | 96.94％ 96.94％ 96.43％ 96.94％ 95.92％ 95.92％ 95.41％ 94.90％ 94.39％ 95.92％ | 88.49％ 87.80％ 87.80％ 85.40％ 84.71％ 83.85％ 84.19％ 82.99％ 81.96％ 80.93％ | 94.36％ 94.07％ 93.99％ 92.89％ 92.38％ 91.94％ 91.72％ 91.14％ 90.62％ 90.40％ | 99.32％ 99.32％ 99.32％ 99.66％ 99.32％ 99.49％ 99.49％ 99.83％ 99.66％ 99.66％ | 96.43％ 96.43％ 95.92％ 96.43％ 95.92％ 95.41％ 95.41％ 95.41％ 94.90％ 96.43％ | 87.63％ 87.80％ 88.32％ 87.29％ 87.11％ 86.94％ 86.94％ 86.43％ 84.88％ 85.05％ | 93.92％ 93.99％ 94.14％ 93.92％ 93.63％ 93.55％ 93.55％ 93.48％ 92.67％ 92.97％ |
It is 70% o'clock in the rate of blocking that table 6 has provided test data, uses I class methods BB type operation simulation shelter target, to tertiary target BMP2 armo(u)red carrier, BTR70 armo(u)red carrier and T72 main battle tank identified obfuscated matrix.
Table 7 has provided that 15 ° of depression angle images are used for doing test data in the MSTAR data, adopt II class methods BB type operation simulation shelter target, the experimental result of using SVM and KNN sorter that tertiary target is discerned is comprising every classification target discrimination and average recognition rate.Fig. 8 is and the corresponding discrimination curve of table 7, Fig. 8 (a) is that the test data rate of blocking was simulated 90% o'clock from 0%, adopt four kinds of average recognition rate curve maps that the contrast sorter is discerned tertiary target, Fig. 8 (b)～(d) is for using four kinds of discrimination curve maps that the contrast sorter is discerned tertiary target respectively.
The confusion matrix of table 6 I class methods, 70% blocks the result's (BB type) under the rate
| 70% test data of blocking | (1) single training set | (2) many training sets | |||||
| BMP2 | BTR70 | T72 | BMP2 | BTR70 | T72 | ||
| KNN | BMP2 BTR70 T72 | 578 9 109 | 5 187 4 | 4 0 469 | 580 8 113 | 4 188 5 | 3 0 464 |
| SVM | BMP2 BTR70 T72 | 575 8 96 | 3 186 3 | 9 2 483 | 586 7 76 | 1 187 3 | 0 2 503 |
The average recognition rate of table 7 II class methods and each identification of targets rate (BB type)
| Sorter | Test data is blocked rate | (1) single training set | (2) many training sets | ||||||
| BMP2 | BTR70 | T72 | Average recognition rate | BMP2 | BTR70 | T72 | Average recognition rate | ||
| KNN | 0 10％ 20％ 30％ 40％ 50％ 60％ 70％ 80％ 90％ | 99.66％ 99.66％ 99.66％ 99.15％ 98.98％ 99.49％ 99.49％ 99.15％ 98.81％ 98.98％ | 96.43％ 95.92％ 95.41％ 94.90％ 94.90％ 94.90％ 94.90％ 95.41％ 96.43％ 93.88％ | 84.19％ 84.02％ 82.82％ 83.68％ 83.33％ 82.82％ 82.13％ 81.62％ 80.93％ 79.55％ | 92.60％ 92.45％ 91.87％ 91.94％ 91.72％ 91.72％ 91.43％ 91.14％ 90.84％ 89.96％ | 99.66％ 99.15％ 99.49％ 99.49％ 99.49％ 99.49％ 99.32％ 99.49％ 98.64％ 98.98％ | 96.43％ 96.43％ 96.43％ 95.92％ 96.43％ 96.43％ 96.43％ 94.90％ 95.92％ 95.41％ | 84.71％ 84.71％ 83.51％ 83.68％ 83.85％ 82.99％ 83.85％ 82.13％ 81.10％ 80.07％ | 92.82％ 92.60％ 92.23％ 92.23％ 92.38％ 92.01％ 92.31％ 91.43％ 90.77％ 90.40％ |
| SVM | 0 10％ 20％ 30％ 40％ 50％ 60％ 70％ 80％ 90％ | 99.32％ 99.32％ 99.66％ 99.32％ 99.15％ 98.81％ 99.15％ 99.49％ 98.98％ 99.49％ | 96.94％ 96.94％ 96.94％ 96.94％ 96.94％ 96.43％ 96.43％ 95.92％ 96.94％ 96.43％ | 88.49％ 88.32％ 88.14％ 87.80％ 86.43％ 85.74％ 85.57％ 84.71％ 85.22％ 83.85％ | 94.36％ 94.29％ 94.36％ 94.07％ 93.41％ 92.89％ 92.97％ 92.67％ 92.89％ 92.38％ | 99.49％ 99.49％ 99.49％ 99.49％ 99.32％ 99.49％ 99.49％ 99.49％ 99.49％ 99.49％ | 96.94％ 96.43％ 96.94％ 96.43％ 96.94％ 96.94％ 96.43％ 96.43％ 95.92％ 96.43％ | 87.11％ 87.46％ 87.63％ 87.97％ 87.80％ 87.29％ 86.08％ 85.91％ 85.05％ 85.57％ | 93.85％ 93.92％ 94.07％ 94.14％ 94.07％ 93.92％ 93.33％ 93.26％ 92.82％ 93.11％ |
It is 70% o'clock in the rate of blocking that table 8 has provided test data, uses II class methods BB type operation simulation shelter target, to tertiary target BMP2 armo(u)red carrier, BTR70 armo(u)red carrier and T72 main battle tank identified obfuscated matrix.
Summary analysis 5, table 6, table 7 and table 8, can see that the BB pattern is intended operation and the operation of OB pattern plan can obtain similar conclusion, be the increase not generation obviously decline of the discrimination of method proposed by the invention along with the rate of blocking, this method that shows that the present invention proposes has excellent performance when height blocks rate.
Result's (BB type) under the confusion matrix .70% shield rate of table 8 II class methods
| 70% test data of blocking | (1) single training set | (2) many training sets | |||||
| BMP2 | BTR70 | T72 | BMP2 | BTR70 | T72 | ||
| KNN | BMP2 BTR70 T72 | 582 8 104 | 3 187 3 | 2 1 475 | 584 9 102 | 3 186 2 | 0 1 478 |
| SVM | BMP2 BTR70 T72 | 584 6 87 | 1 188 2 | 2 2 493 | 584 5 80 | 1 189 2 | 2 2 500 |
By Fig. 7 and Fig. 8 as seen, for SVM and two kinds of sorters of KNN, use single training set and many training sets training classifier not to have significant difference.For KNN, compare with single training set, many training sets have increased the shelter target image, because after former target peak point is blocked, the position and the amplitude of the noise spot that adds are at random, big variation does not take place in the whole pixel value of shelter target image, the number of peak value unique point is counted out with " bright " in the target zone and is compared still seldom in addition, so the classification situation that the test pattern neighbour is ordered can not produce great changes along with the adding of shielded image, thereby many training sets have increased shelter target and do not make the discrimination of KNN produce great changes.For SVM, many training set training are significantly improved than the discrimination of single training set training, illustrate that the overall distribution influence of blocking image is very big.Because during single training set training classifier, the target of test pattern added block after, variation has taken place in the characteristic of target image and the former target image that do not have when blocking is disobeyed same distribution, so discrimination is lower; And during many training sets training classifier, added shelter target in the training image after, just comprise the target of obeying same distribution with the test sample book target in the training set, so discrimination is significantly increased.
Fig. 9 (a) has provided when adopting operation of I class methods BB type and I class methods OB type operation simulation shelter target respectively, use the svm classifier device of single training set and many training sets training, and set 30 scattering centers and set the Bhanu method of 50 scattering centers, the average recognition rate curve map that tertiary target BMP2 armo(u)red carrier, BTR70 armo(u)red carrier and T72 main battle tank are discerned.Fig. 9 (b) has provided when adopting operation of II class methods BB type and II class methods OB type operation simulation shelter target respectively, use the svm classifier device of single training set and many training sets training, and set the Bhanu method of 30 scattering centers and 50 scattering centers, the average recognition rate curve map that tertiary target BMP2 armo(u)red carrier, BTR70 armo(u)red carrier and T72 main battle tank are discerned.As can see from Figure 9, be higher than at 80% o'clock in the rate of blocking, the discrimination of I class methods and II class methods all is higher than the Bhanu method, the relative OB type operation of the recognition result that the operation of BB type obtains is milder, not obviously decline of generation of increase discrimination along with the rate of blocking shows the excellent properties of method when height blocks rate that the present invention proposes.
Claims (5)
1. based on the method for synthetic aperture radar shelter from shelter from object identification of supporting vector machine, the specific implementation process is as follows:
(1), the SAR target image to be identified of input is carried out pre-service, obtain the ROI of image after the pre-service;
(2), after pre-service, seek the peak point of target among the ROI of image;
(3), to pretreated ROI, by removing the part peak point of target, the peak point that adds respective number with I class or II class methods is simulated shelter target in ROI;
(4), select sorter for use and sorter trained, in the sorter that the test pattern input is trained, carry out shelter target identification.
2. the SAR target image to be identified to input according to claim 1 carries out pre-service, obtains the ROI of image after the pre-service, and its method is as follows:
(1), with median filter to the input picture denoising, to weaken the influence of speckle noise;
(2), image is carried out normalization, image pixel value is normalized to 0～1;
(3), image after the normalization is carried out Threshold Segmentation, the threshold value of cutting apart is the standard deviation that the average of image adds twice after the normalization, is generally 0.8;
(4), image after the Threshold Segmentation is carried out the mathematical morphology conversion: the method that employing is expanded makes the interregional slit that splits be filled, and corrodes to make the zone become the target main areas of image and several zonules on every side again;
(5), the wavenumber filter of using tricks carries out filtering, setting counting wave filter size is 5 * 5, is in the image-region of 5 * 5 sizes at center with certain pixel promptly, if pixel value be zero number of pixels greater than 5 * 5/2, then this pixel value is adjusted into zero, otherwise, do not adjust;
(6), carry out expansive working, remedy some bright spots that the target area edge penalty falls, promptly at the non-vanishing point of some pixel values of target area edge supplement;
(7), find out the position of the non-vanishing point of the pixel value of (6) resultant image, to in the input picture pixel that should the position be kept initial value, the pixel value zero setting of rest of pixels, taking out size from the input picture center position then is 54 * 54 subgraph, has just obtained the ROI of input picture.
3. the method for synthetic aperture radar shelter from shelter from object identification based on supporting vector machine according to claim 1, the said peak point of after pre-service, seeking target among the ROI of image, its method is to search for successively according to raster scan order, if this point is to be the point that has max pixel value in 3 * 3 image-regions at center with it, then this is a peak point, writes down its pixel value and coordinate; Otherwise continue search.
4. the method for synthetic aperture radar shelter from shelter from object identification based on supporting vector machine according to claim 1, said to pretreated ROI, by removing the part peak point of target, add the peak point of respective number in ROI with I class or II class methods, simulate shelter target, its process is as follows:
(1), determine after the target peak point of ROI, from four direction: against the orientation of aircraft to d
1, go up distance to d
2, along the orientation of aircraft to d
3With following distance to d
4In either direction begin, remove and run into successively, number is the peak point of peak point sum N%, simulates the shelter target that the rate of blocking is N%, N ∈ [0,100];
(2), the random site that the pixel of number same as described above, random value is added the zone that is deleted peak point.In the present invention, adopt I class or two kinds of methods of II class to simulate: the pixel that adds in the I class methods is obeyed the even distribution in [0,1]; The pixel that adds in the II class methods is obeyed the even distribution in [0.8,1].
5. the method for synthetic aperture radar shelter from shelter from object identification based on supporting vector machine according to claim 1, said (4), select sorter for use and sorter trained, in the sorter that test pattern input is trained, carry out shelter target identification, wherein:
Saidly select sorter for use and sorter is trained, its implementation procedure is: on a surface target imaging data was as training in the training sample input supporting vector machine when MSTAR was 17 ° in the depression angle, whole training datas are divided into one group according to its azimuth information since 0 ° of per 10 ° of interval, being divided into is 36 groups, accordingly 36 sorters is trained; According to the training data difference, adopt following two kinds of methods to train:
2., adopt single training set, promptly include only unscreened image to be identified in the training set;
2., adopt many training sets, promptly both comprised unscreened image to be identified in the training set, comprise again with image to be identified and simulate the shelter target image of the rate of blocking under 10%～90% situation;
In the said sorter that test pattern input is trained, discern, its implementation procedure is: on a surface target imaging data was as test pattern when MSTAR was 15 ° in the depression angle, and, just can finish identification to shelter target with in the above-mentioned counterparty's parallactic angle sorter that has trained of this test pattern input.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CNB2007100185741A CN100573190C (en) | 2007-09-04 | 2007-09-04 | Method for synthetic aperture radar shelter from shelter from object identification based on supporting vector machine |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CNB2007100185741A CN100573190C (en) | 2007-09-04 | 2007-09-04 | Method for synthetic aperture radar shelter from shelter from object identification based on supporting vector machine |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN101135729A true CN101135729A (en) | 2008-03-05 |
| CN100573190C CN100573190C (en) | 2009-12-23 |
Family
ID=39159910
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CNB2007100185741A Expired - Fee Related CN100573190C (en) | 2007-09-04 | 2007-09-04 | Method for synthetic aperture radar shelter from shelter from object identification based on supporting vector machine |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN100573190C (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101807258B (en) * | 2010-01-08 | 2012-05-23 | 西安电子科技大学 | SAR (Synthetic Aperture Radar) image target recognizing method based on nuclear scale tangent dimensionality reduction |
| CN104614722A (en) * | 2015-01-20 | 2015-05-13 | 南京大学 | Method for identifying radar shielding based on signal-to-noise ratio |
| CN105403872A (en) * | 2015-11-04 | 2016-03-16 | 北京无线电测量研究所 | Moving vehicle target classification method and system |
| CN107728115A (en) * | 2017-09-11 | 2018-02-23 | 电子科技大学 | Ambient interferences suppressing method based on SVM after a kind of radar target imaging |
| CN108510467A (en) * | 2018-03-28 | 2018-09-07 | 西安电子科技大学 | SAR image target recognition method based on variable depth shape convolutional neural networks |
| CN110335285A (en) * | 2019-07-08 | 2019-10-15 | 中国科学院自动化研究所 | SAR image target label method, system, device based on rarefaction representation |
| CN110798677A (en) * | 2018-08-01 | 2020-02-14 | Oppo广东移动通信有限公司 | Three-dimensional scene modeling method and device, electronic device, readable storage medium and computer equipment |
| CN111650564A (en) * | 2020-02-26 | 2020-09-11 | 惠州市德赛西威智能交通技术研究院有限公司 | Vehicle-mounted radar failure detection method based on CFAR post-peak point statistics |
| CN111870279A (en) * | 2020-07-31 | 2020-11-03 | 西安电子科技大学 | Method, system and application for segmenting left ventricular myocardium of ultrasonic image |
| CN114267076A (en) * | 2022-03-01 | 2022-04-01 | 天津所托瑞安汽车科技有限公司 | Image identification method, device, equipment and storage medium |
-
2007
- 2007-09-04 CN CNB2007100185741A patent/CN100573190C/en not_active Expired - Fee Related
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101807258B (en) * | 2010-01-08 | 2012-05-23 | 西安电子科技大学 | SAR (Synthetic Aperture Radar) image target recognizing method based on nuclear scale tangent dimensionality reduction |
| CN104614722A (en) * | 2015-01-20 | 2015-05-13 | 南京大学 | Method for identifying radar shielding based on signal-to-noise ratio |
| CN104614722B (en) * | 2015-01-20 | 2018-05-08 | 南京大学 | A kind of method blocked based on signal-to-noise ratio Discrimination Radar |
| CN105403872A (en) * | 2015-11-04 | 2016-03-16 | 北京无线电测量研究所 | Moving vehicle target classification method and system |
| CN107728115A (en) * | 2017-09-11 | 2018-02-23 | 电子科技大学 | Ambient interferences suppressing method based on SVM after a kind of radar target imaging |
| CN107728115B (en) * | 2017-09-11 | 2020-08-11 | 电子科技大学 | SVM-based background interference suppression method after radar target imaging |
| CN108510467A (en) * | 2018-03-28 | 2018-09-07 | 西安电子科技大学 | SAR image target recognition method based on variable depth shape convolutional neural networks |
| CN108510467B (en) * | 2018-03-28 | 2022-04-08 | 西安电子科技大学 | SAR image target identification method based on depth deformable convolution neural network |
| CN110798677B (en) * | 2018-08-01 | 2021-08-31 | Oppo广东移动通信有限公司 | Three-dimensional scene modeling method and device, electronic device, readable storage medium and computer equipment |
| CN110798677A (en) * | 2018-08-01 | 2020-02-14 | Oppo广东移动通信有限公司 | Three-dimensional scene modeling method and device, electronic device, readable storage medium and computer equipment |
| CN110335285A (en) * | 2019-07-08 | 2019-10-15 | 中国科学院自动化研究所 | SAR image target label method, system, device based on rarefaction representation |
| CN110335285B (en) * | 2019-07-08 | 2022-04-26 | 中国科学院自动化研究所 | SAR image target marking method, system and device based on sparse representation |
| CN111650564A (en) * | 2020-02-26 | 2020-09-11 | 惠州市德赛西威智能交通技术研究院有限公司 | Vehicle-mounted radar failure detection method based on CFAR post-peak point statistics |
| CN111870279A (en) * | 2020-07-31 | 2020-11-03 | 西安电子科技大学 | Method, system and application for segmenting left ventricular myocardium of ultrasonic image |
| CN114267076A (en) * | 2022-03-01 | 2022-04-01 | 天津所托瑞安汽车科技有限公司 | Image identification method, device, equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN100573190C (en) | 2009-12-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN100573190C (en) | Method for synthetic aperture radar shelter from shelter from object identification based on supporting vector machine | |
| CN108510467B (en) | SAR image target identification method based on depth deformable convolution neural network | |
| CN101976437B (en) | High-resolution remote sensing image variation detection method based on self-adaptive threshold division | |
| CN107067415B (en) | A kind of object localization method based on images match | |
| CN109766835B (en) | SAR target recognition method for generating countermeasure network based on multi-parameter optimization | |
| CN102214298B (en) | Method for detecting and identifying airport target by using remote sensing image based on selective visual attention mechanism | |
| DE60038158T2 (en) | OBJECT DETECTION APPARATUS AND METHOD FOR ESTIMATING THE AZIMUM OF RADAR TARGETS BY RADON TRANSFORMATION | |
| Liu et al. | Automated vehicle extraction and speed determination from QuickBird satellite images | |
| CN104851097B (en) | The multichannel SAR GMTI methods aided in based on target shape and shade | |
| CN102844768B (en) | The shielding of image template | |
| CN103218831A (en) | Video moving target classification and identification method based on outline constraint | |
| CN104951799B (en) | A kind of SAR remote sensing image oil spilling detection recognition method | |
| CN101964060B (en) | SAR variant target identification method based on local textural feature | |
| CN109635789B (en) | High-resolution SAR image classification method based on intensity ratio and spatial structure feature extraction | |
| CN101839722A (en) | Method for automatically recognizing target at medium and low altitudes and positioning carrier with high accuracy | |
| CN107103338A (en) | Merge the SAR target identification methods of convolution feature and the integrated learning machine that transfinites | |
| CN110147598B (en) | Ultrahigh-speed impact fragment cloud modeling and damage evaluation method based on image processing | |
| CN107230203A (en) | Casting defect recognition methods based on human eye vision attention mechanism | |
| CN103824088A (en) | SAR target variant recognition method based on multi-information joint dynamic sparse representation | |
| CN102385690A (en) | Target tracking method and system based on video image | |
| CN102609701A (en) | Remote sensing detection method based on optimal scale for high-resolution SAR (synthetic aperture radar) | |
| CN103020649A (en) | Forest type identification method based on texture information | |
| CN109886267A (en) | A kind of soft image conspicuousness detection method based on optimal feature selection | |
| CN101515369A (en) | Multiscale SAR image segmentation method based on semi-supervised learning | |
| CN109101926A (en) | Aerial target detection method based on convolutional neural networks |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20091223 Termination date: 20200904 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |