CN106599831B - Based on the specific SAR target discrimination method with shared dictionary of sample weighting classification - Google Patents

Abstract

The invention discloses a kind of based on the specific SAR target discrimination method with shared dictionary of sample weighting classification, and it is low mainly to solve the problems, such as that prior art SAR target under complex scene identifies performance.Its scheme is: 1. pairs of given training slices and test slice extract local feature；2. obtaining Global Dictionary using the local feature of training slice；3. being sliced using Global Dictionary to training and the local feature of test slice carrying out standardized sparse coding respectively, local feature code coefficient is obtained；4. the local feature code coefficient of pair training slice and test slice carries out feature merging and dimensionality reduction respectively, the global characteristics of training slice and the global characteristics of test slice are obtained；5. being identified using support vector machines to test slice global characteristics.The present invention improves the performance of identification, can be used for identifying the SAR target of complex scene.

Description

Based on the specific SAR target discrimination method with shared dictionary of sample weighting classification

Technical field

The invention belongs to radar target authentication technique fields, relate generally to a kind of SAR target discrimination method, can be used for vehicle Target recognition and classification provides important information.

Background technique

Synthetic aperture radar SAR utilizes microwave remote sensing technique, climate and does not influence round the clock, with round-the-clock, round-the-clock Ability to work, and have the characteristics that multiband, multipolarization, visual angle be variable and penetrability.With more and more airborne and stars The appearance for carrying SAR, brings the SAR data under a large amount of different scenes, is exactly that automatic target is known to SAR data one important application Other ATR.Target under complex scene, which identifies, also becomes one of current research direction.

Feature extraction is an important process in target discrimination process.In in the past few decades, have largely about The research that SAR target diagnostic characteristics extract, be broadly divided into four kinds: the first is characterized in textural characteristics, as Lincoln laboratory proposes Standard deviation characteristic, FRACTAL DIMENSION feature and arrangement energy ratio feature；Second of feature is related with the shape of target, such as ERIM Qualitative character, characteristics of diameters and the variance that (Environmental Research Institute of Michigan) is proposed are returned One changes horizontal and vertical projection properties used in rotary inertia feature and other documents, minimum and maximum projected length spy Sign；The third feature is decided by the contrast of target and background, the peak C FAR and mean value CFAR and CFAR proposed such as ERIM The average signal-to-noise ratio that most bright percentage feature and Gao are proposed, Y-PSNR and brightest pixel percentage feature.In addition to this, Lincoln laboratory also proposed what high luminance pixel when several features are used to describe to image plus different threshold values was spread in space Change, the difference that these features depend not only on target and background additionally depends on the size of target；4th kind of feature is that polarization is special Sign, such as percent purity, pure idol percentage and strong even percentage feature, but these polarization characteristics can only be in full-polarization SAR number It can be just extracted according to upper.

The shortcomings that above-mentioned traditional characteristic mainly has in terms of following two: first, these features target is only provided it is coarse, Partial description, they cannot describe target and the detailed local shape of clutter and structural information, this shows that identification cannot be abundant Utilize full resolution pricture detailed information abundant.When target and clutter are no apparent poor in terms of texture, shape and contrast When other, these features cannot show to identify performance well.Second, existing feature is suitable for naturally miscellaneous under simple scenario The identification of wave and target.The verifying of current most of SAR target discrimination methods is all based on MSTAR data set, with 0.3m points Resolution.The scene of this standard data set is fairly simple, and target slice possesses similar feature, and each slice only includes a mesh Mark and the center for being located at sectioning image.Target is a compact high intensity region, is around that intensity is lower, background of homogeneity Clutter.Clutter slice also shows some similar attributes, and most of high-intensitive region corresponds to tree crown in clutter slice.These Target slice and clutter slice differ greatly on texture, shape and contrast, and traditional target diagnostic characteristics are suitable for the number According to collection, and show relatively good identification feature.However, true scene is more complicated, such as miniSAR data set, target The position and direction of target are different in slice, and are had existing for multiple target or partial target in a width sectioning image Situation.Clutter is sliced, the type of clutter is diversified, including natural clutter, and such as trees, there are many more artificial miscellaneous Wave, such as the edge of building.Therefore existing texture, shape and contrast metric be not enough to identify target in this case and Clutter.

In conclusion traditional characteristic identifies tool to the target under complex scene with the continuous promotion of SAR image resolution ratio There is biggish limitation.

Summary of the invention

It is an object of the invention to the deficiencies for existing SAR target discrimination method, propose a kind of based on sample weighting class The not specific SAR target discrimination method with shared dictionary identifies performance with the target improved under complex scene.

The technical scheme of the present invention is realized as follows:

(1) using SAR-SIFT descriptor to given training sectioning imageIt is sliced with test ImageLocal feature is extracted, the local feature for training sectioning image is obtainedWith the local feature of test sectioning imageWherein,Indicate clutter class training sectioning image,Indicate target class training sectioning image,Indicate clutter Class testing sectioning image,Indicate target class testing sectioning image,It is clutter class training sectioning image Local feature,It is the local feature of target class training sectioning image,It is clutter class testing slice The local feature of image,It is the local feature of target class testing sectioning image, p₁Indicate clutter class training slice map As number, p₂Indicate target class training sectioning image number, k₁Indicate clutter class testing sectioning image number, k₂Indicate target class Test sectioning image number；

(2) by the clutter class training sectioning image local feature in (1) resulting XAs the training of clutter class Sample, target class training sectioning image local featureAs target class training sample, Global Dictionary U is obtained；

2a) initialize clutter category dictionary U₁, target category dictionary U₂, shared dictionary U₀, clutter class training sample weight With target class training sample weightIf current iteration number iter=0；

2b) according to the clutter category dictionary U under current iteration number₁, target category dictionary U₂With shared dictionary U₀, calculate clutter Class training slice local featureRarefaction representation coefficient H₁With target class training slice local featureRarefaction representation coefficient H₂；

2c) according to 2b) obtained H₁And H₂, using alternate optimization method, update clutter category dictionary U₁, target category dictionary U₂ With shared dictionary U₀, obtain updated clutter category dictionary U₁', target category dictionary U₂' and shared dictionary U₀′；

Iter=iter+1 2d) is enabled, current the number of iterations is recorded, sample weights update is judged whether to, if mod (iter, iterSkip) is equal to 0, executes step 2e) it is trained sample weights update；Otherwise, without training sample weight It updates, enables U₁=U₁′、U₂=U₂′、U₀=U₀' return step 2b), wherein iterSkip indicates that training sample weight updates interval, Mod expression takes the remainder；

2e) utilize 2c) obtain U₁′、U₂' and U₀' update clutter class training sample weightObtain updated clutter Class training sample weightUtilize 2c) obtain U₁′、U₂' and U₀' update target class training sample weightIt obtains Updated target class training sample weight

2f) judge whether current iteration number iter is less than maximum number of iterations iterMax, if being less than, enables U₁=U₁′、 U₂=U₂′、U₀=U₀′、Return step 2b), if being equal to, iteration stopping, Obtain final Global Dictionary U=[U₀′,U₁′,U₂′]；

(3) the Global Dictionary U for utilizing (2) to obtain obtains training the local feature X of sectioning image to (1) and test is sliced The local feature Y of image carries out standardized sparse coding respectively, obtains the local feature code coefficient for training sectioning imageWith test sectioning image local feature code coefficient:

(4) local feature of the local feature code coefficient V for the training sectioning image for obtaining (3) and test sectioning image Code coefficient W carries out feature merging and dimensionality reduction respectively, obtained training sectioning image global characteristics:

With the global characteristics of test sectioning image

(5) using global characteristics V " ' one two class Linear SVM classifier of training of training sectioning image, using training Classifier to test sectioning image global characteristics W " ' classify, obtain it is each test sectioning image categorised decision value The categorised decision value decision is compared with the threshold value Thr=0 of setting, if decision >=Thr, recognizes by decision It is otherwise clutter class slice to be target class slice.

Compared with the prior art, the present invention has the following advantages:

1. the present invention is the SAR image vehicle target discrimination method under complex scene, the target compared to traditional characteristic is reflected Other method, the distribution due to considering target and the partial structurtes information and partial structurtes of clutter slice under complex scene are believed Breath, takes full advantage of the detailed information of full resolution pricture, and the SAR target improved under complex scene identifies performance.

It is and existing 2. the present invention is during generating global dictionary due to increasing the study to descriptive bad sample Based on the specific target class global characteristics compared with the SAR target discrimination method of shared dictionary learning CSDL, obtained of classification with The discrimination of the global characteristics of clutter class is bigger, to further improve the identification performance of the SAR target under complex scene.

Detailed description of the invention

Fig. 1 is implementation flow chart of the invention；

Fig. 2 is that the Global Dictionary in the present invention generates sub-process figure；

Fig. 3 is part miniSAR sectioning image used in present invention experiment 1；

Fig. 4 is part miniSAR sectioning image used in present invention experiment 2；

Fig. 5 is part miniSAR sectioning image used in present invention experiment 3；

Fig. 6 is part miniSAR sectioning image used in present invention experiment 4.

Specific embodiment

The embodiment of the present invention and effect are described in further detail with reference to the accompanying drawing:

The method of the present invention relates generally to the identification of the vehicle target under complex scene, and existing target diagnostic characteristics are mostly It is verified based on MSTAR data set, the scene of data description is relatively simple.Target slice possesses similar feature, each Slice is only comprising a target and positioned at the center of sectioning image.Target area is compact a, high intensity region, surrounding It is that intensity is lower, clutter background of homogeneity.Clutter slice also shows some similar attributes, most of height in clutter slice The region of intensity corresponds to tree crown.These target slices and clutter slice differ greatly on texture, shape and contrast.With thunder Up to the promotion of resolution ratio, the scene of SAR image description is also increasingly complex, and target slice not only has single goal, and there are also multiple target drawn games The case where portion's target, and target is also not necessarily located in the center of slice.Clutter slice is also not only nature clutter, and there are also a large amount of shapes The different artificial clutter of shape.In view of the above problems, to be taken based on sample weighting classification specific with shared dictionary learning phase by the present invention In conjunction with, SAR target is identified, improve under complex scene to the identification performance of SAR target.

Referring to Fig. 1, realization step of the invention includes the following:

Step 1, local feature is extracted to given training sectioning image and test sectioning image.

2a) using SAR-SIFT descriptor to given training sectioning imageIt is special to carry out part Sign is extracted, and the local feature for training sectioning image is obtainedWhereinIndicate clutter Class trains sectioning image,Indicate target class training sectioning image,It is clutter class training sectioning image Local feature,It is the local feature of target class training sectioning image, p₁Indicate clutter class training sectioning image Number, p₂Indicate target class training sectioning image number；

2b) using SAR-SIFT descriptor to given test sectioning imageCarry out part Feature extraction obtains the local feature of test sectioning imageWherein,Indicate clutter Class testing sectioning image,Indicate target class testing sectioning image,It is clutter class testing sectioning image Local feature,It is the local feature that target class testing practices sectioning image, k₁Indicate clutter class testing sectioning image number Mesh, k₂Indicate target class testing sectioning image number.

Step 2, according to the local feature of training sectioning imageObtain Global Dictionary U.

By the local feature of clutter class training sectioning imageAs clutter class training sample, target class is instructed Practice the local feature of sectioning imageAs target class training sample, Global Dictionary U is obtained.

Referring to Fig. 2, this step is implemented as follows:

2a) initialize clutter category dictionary U₁, target category dictionary U₂, shared dictionary U₀, clutter class training sample weight With target class training sample weight

2a1) from10000 local features are randomly selected, with K-SVD algorithm to clutter category dictionaryInitialization, with Lagrange duality algorithm by U₁It updates once, wherein d indicates training sectioning image local feature Dimension, n₁Indicate clutter category dictionary atom number；

2a2) from10000 local features are randomly selected, with K-SVD algorithm to target category dictionaryInitialization, with Lagrange duality algorithm by U₂It updates once, wherein n₂Indicate target category dictionary atom number；

2a3) fromWithIn randomly select 10000 local features, with K-SVD algorithm to altogether Enjoy dictionaryInitialization, with Lagrange duality algorithm by U₀It updates once, wherein n₀Indicate shared dictionary atom Number；

2a4) by clutter class training sample weightWith target class training sample weightIt is initialized to 1；

2a5) set current iteration number iter=0；

2b) according to the clutter category dictionary U under current iteration number₁, target category dictionary U₂With shared dictionary U₀, calculate clutter Class training slice local featureRarefaction representation coefficient H₁With target class training slice local featureRarefaction representation coefficient H₂, steps are as follows for calculating:

Following optimization problems 2b1) are solved by feature-sign search algorithm, i-th of clutter class training is obtained and cuts The local feature of pictureRarefaction representation coefficient

Wherein i=1 ..., p₁, λ expression weighting parameters, | | | |_FIndicate F norm, | | | |₁Indicate l₁Norm,

The local feature of all clutter class training sectioning images is solvedAfter rarefaction representation coefficient, obtain more The local feature rarefaction representation coefficient of clutter class training sectioning image after new

Following optimization problems 2b2) are solved by feature-sign search algorithm, j-th of target class training is obtained and cuts The local feature of pictureRarefaction representation it is sparse

Wherein j=1 ..., p₂,

The local feature of all target class training sectioning images is solvedAfter rarefaction representation coefficient, obtain more The local feature rarefaction representation coefficient of target class training sectioning image after new

2c) according to 2b) obtained H₁And H₂, using alternate optimization method, update clutter category dictionary U₁, target category dictionary U₂ With shared dictionary U₀, steps are as follows for update:

Following optimization problems 2c1) are solved by alternate optimization method, update clutter class U₁Dictionary obtains updated miscellaneous Wave category dictionary U₁':

s.t.||U₁(:,b₁)||₂=1, b₁=1 ..., n₁

Wherein, η₁₁WithIt is weighting parameters, | | | |₂It is l₂Norm, | | | |_FIt is F norm,Be size be n₀List Bit matrix,Be size be n₁×n₀Null matrix,Be size be n₁Unit matrix,Be size be n₀×n₁'s Null matrix, n₁It is clutter category dictionary U₁Atomicity, n₂It is target category dictionary U₂Atomicity, n₀It is shared dictionary U₀Atom Number, n=n₀+n₁+n₂, W₁It is clutter class training sample weight matrix:

m₁=n_L×p₁It is the local feature sum of clutter class training sectioning image, n_LIt indicates in a trained sectioning image Local feature number.

Following optimization problems 2c2) are solved by alternate optimization method, update target category dictionary U₂, obtain updated mesh Mark category dictionary U₂':

s.t.||U₂(:,b₂)||₂=1, b₂=1 ..., n₂

Wherein, η₂₁WithIt is weighting parameters, | | | |₂It is l₂Norm, | | | |_FIt is F norm,Be size be n₀'s Unit matrix,Be size be n₂×n₀Null matrix,Be size be n₂Unit matrix,Be size be n₀×n₂ Null matrix, n₁It is clutter category dictionary U₁Atomicity, n₂It is target category dictionary U₂Atomicity, n₀It is shared dictionary U₀Atom Number, n=n₀+n₁+n₂, W₂It is target class training sample weight matrix:

m₂=n_L×p₂It is the local feature sum of target class training sectioning image, n_LIt indicates in a trained sectioning image Local feature number.

Following optimization problems 2c3) are solved by alternate optimization method, update shared dictionary U₀, obtain updated shared Dictionary U₀':

s.t.||U₀(:,b₀)||₂,b₀=1 ..., n₀

Wherein, η₀₁WithIt is weighting parameters, | | | |₂It is l₂Norm, | | | |_FIt is F norm,n₁It is Clutter category dictionary U₁Atomicity, n₂It is target category dictionary U₂Atomicity, n₀It is shared dictionary U₀Atomicity, n=n₀+n₁+ n₂, It is size For n₀Unit matrix,Be size be n₁×n₀Null matrix,Be size be n₁Unit matrix,Be size be n₀ ×n₁Null matrix, W₁It is clutter class training sample weight matrix:

m₁=n_L×p₁It is the local feature sum of clutter class training sectioning image, n_LIt indicates in a trained sectioning image Local feature number, Be size be n₀Unit matrix,Be size be n₂×n₀Null matrix,Be size be n₂Unit matrix,It is Size is n₀×n₂Null matrix, W₂It is target class training sample weight matrix:

m₂=n_L×p₂It is the local feature sum of target class training sectioning image.

After completing above-mentioned update step, updated clutter category dictionary U is obtained₁', target category dictionary U₂', shared dictionary U₀′；

Iter=iter+1 2d) is enabled, current the number of iterations is recorded, sample weights update is judged whether to, if mod (iter, iterSkip) is equal to 0, executes step 2e) it is trained sample weights update；Otherwise, without training sample weight It updates, enables U₁=U₁′、U₂=U₂′、U₀=U₀' return step 2b), wherein iterSkip indicates that training sample weight updates interval, Mod expression takes the remainder；

2e) utilize 2c) obtain U₁′、U₂' and U₀' update clutter class training sample weightWith target class training sample WeightIts step are as follows,

2e1) utilize 2c) obtain U₁′、U₂' and U₀' update clutter class training sample weightIt obtains updated miscellaneous The weight of wave class training sampleWherein i-th of clutter class training sample weight w_1i' obtained by following equations,

In formula, i=1 ..., p₁, α is a scaling factor bigger than 1, w_mIt is the maximum in weight allowed band Value,It is the local feature X of i-th of clutter class training sectioning image₁ ⁱRarefaction representation coefficient, value utilize feature- Sign search algorithm solving optimization problemIt obtains,It is U₀' corresponding Rarefaction representation coefficient,It is U₁' corresponding rarefaction representation coefficient,It is U₂' corresponding rarefaction representation coefficient, It is clutter sectioning image local featureUse target category dictionary U₂The average energy of ' reconstruct；2e2) utilize 2c) obtain U₁′、U₂′ And U₀' update target class training sample weightObtain the weight of updated target class training sampleIts In j-th of target class training sample weight w_2j' obtained by following equations,

In formula, j=1 ..., p₂,It is the local feature X of j-th of target class training sectioning image₂ ^jRarefaction representation system Number, value utilize feature-sign search algorithm solving optimization problem:

It obtains,It is U₀' corresponding rarefaction representation coefficient,It is U₁' corresponding rarefaction representation coefficient,It is U₂' right The rarefaction representation coefficient answered,It is target slice image local featureUse clutter category dictionary U₁' reconstruct is put down Equal energy；

2f) judge whether current iteration number iter is less than maximum number of iterations iterMax, if being less than, enables U₁=U₁′、 U₂=U₂′、U₀=U₀′、Return step 2b), if being equal to, iteration stopping, Obtain final Global Dictionary U=[U₀′,U₁′,U₂′]；

Step 3, the local feature of training sectioning image and the local feature code coefficient of test sectioning image are solved.

This step is implemented as follows:

3a) the Global Dictionary U obtained using step 2 obtains step 1 the local feature X of sectioning image is trained to mark Quasi- sparse coding obtains the local feature code coefficient for training sectioning image

3b) the Global Dictionary U obtained using step 2, the local feature Y for obtaining test sectioning image to step 1 are marked Quasi- sparse coding obtains the local feature code coefficient of test sectioning image

Step 4, the local feature code coefficient V of training sectioning image step 3 obtained and the office of test sectioning image Portion feature coding coefficient W carries out feature merging and dimensionality reduction respectively.

4a) using spatial pyramid Matching Model by training sectioning image be divided into size be 1 × 1,2 × 2,4 × 4 three Sub-regions A1, A2, A3；

4b) merged using maximum value by subregion A1, the local feature code coefficient V of the corresponding training sectioning image of A2, A3 It merges and splices, form the global feature of training sectioning image:

To the global feature V ' carry out l of training sectioning image₂Norm normalization, the training slice map after being normalized The global feature V " of picture, wherein h indicates the dimension of global characteristics after feature merging；

Dimensionality reduction, the global characteristics of the training sectioning image after obtaining dimensionality reduction 4c) are carried out to V " using principal component analysisWherein h ' is the dimension of the global characteristics after dimensionality reduction；

4d) using spatial pyramid Matching Model will test sectioning image be divided into size be 1 × 1,2 × 2,4 × 4 this three Sub-regions B1, B2, B3；

4e) merged using maximum value by subregion B1, the local feature code coefficient of the corresponding test sectioning image of B2, B3 W is merged and is spliced, and forms the global feature of test sectioning image:

To the global feature W ' carry out l of test sectioning image₂Norm normalization, the test slice map after being normalized The global characteristics W " of picture；

Dimensionality reduction 4f) is carried out to W " using principal component analysis, the global characteristics of the test sectioning image after obtaining dimensionality reductionWherein h ' is the dimension of the global characteristics after dimensionality reduction.

Step 5, using global characteristics V " ' one two class Linear SVM classifier of training of training sectioning image, training is used Good classifier classifies to the global characteristics W " ' of test sectioning image, obtains the categorised decision of each test sectioning image The categorised decision value decision is compared by value decision with the threshold value Thr=0 of setting, if decision >=Thr, It is considered that target class is sliced, is otherwise clutter class slice.

Effect of the invention can be further illustrated by following experimental data:

Experiment 1:

1.1) experiment scene:

This experiment sectioning image used miniSAR data set disclosed in the U.S. laboratory Sandia, these numbers The website in the laboratory Sandia is downloaded under, partially sliced example images are as shown in figure 3, Fig. 3 (a) is target class training slice map As example, Fig. 3 (b) is clutter class sectioning image example, and Fig. 3 (c) is test sectioning image example.

1.2) four groups of traditional characteristics of experimental selection:

First group of feature is: optimal threshold feature, the average value tag of image pixel quality, image pixel spatial cohesion The combination of feature, corner feature, acceleration signature；

Second group of feature is: the average value tag of optimal threshold feature, image pixel quality, image pixel spatial cohesion Feature, corner feature, acceleration signature, mean value signal-to-noise ratio feature, Y-PSNR feature and brightest pixel percentage feature Combination is used；

Third group feature is: standard deviation characteristic, FRACTAL DIMENSION feature and the combination for arranging energy ratio feature；

4th group of feature is: standard deviation characteristic, FRACTAL DIMENSION feature, arrangement energy ratio feature, optimal threshold feature, image slices The average value tag of quality amount, image pixel spatial cohesion feature, corner feature, acceleration signature, mean value signal-to-noise ratio feature, The combination of Y-PSNR feature and brightest pixel percentage feature.

1.3) experiment parameter:

Training clutter number of slices p₁=1442, training objective number of slices p₂=2091, test clutter number of slices k₁=599, it surveys Try target slice number k₂=140, weighting parameters λ=0.1, scale factor, α=50, weighting parameters η₀₁=η₁₁=η₂₁=η₀₂ =η₁₂=η₂₂=0.05, dictionary learning the number of iterations iterMax=15, sample weights update interval iterSkip=5, dictionary Atomicity n₀=n₁=n₂=300, weight limit value w_m=50, SVM classifier uses LIBSVM kit, SVM punishment in experiment Coefficient C=10；

1.4) experiment content:

With the existing SAR target discrimination method based on first group of traditional characteristic Verbout and the method for the present invention to complexity SAR target under scene compares experiment；

With the existing SAR target discrimination method based on second group of traditional characteristic Verbout+Gao and the method for the present invention pair SAR target under complex scene compares experiment；

With the existing SAR target discrimination method based on third group traditional characteristic Lincoln and the method for the present invention to complexity SAR target under scene compares experiment；

With the existing SAR target discrimination method based on the 4th group of traditional characteristic Lincoln+Verbout+Gao and this hair Bright method compares experiment to the SAR target under complex scene；

With the existing SAR target discrimination method based on CSDL and the method for the present invention to the SAR target under complex scene into Row comparative experiments.

The identification result of experiment 1 is as shown in table 1:

The identification result of 1 distinct methods of table

Distinct methods	AUC	Pc (Thr=0)	Pd (Thr=0)	Pf (Thr=0)	Pd (Thr corresponds to Pd=0.9)	Pf (Thr corresponds to Pd=0.9)
							Verbout	0.8739	87.0095%	0.6143	0.0701	0.9000	0.4040
Verbout+Gao	0.8813	86.1976%	0.6071	0.0785	0.9000	0.3539
							Lincoln	0.9398	90.6631%	0.9571	0.1052	0.9000	0.0801
Lincoln+Verbout+Gao	0.9408	90.3924%	0.9143	0.0985	0.9000	0.0851
							CSDL	0.9580	92.0162%	0.7500	0.0401	0.9000	0.1185
The present invention	0.9694	93.3694%	0.7429	0.0217	0.9000	0.0801

AUC in table 1 indicates that the area under ROC curve, Pc indicate that overall accuracy, Pd indicate that verification and measurement ratio, Pf indicate false-alarm Rate, Thr are the threshold values of SVM classifier.

It can be seen in table 1 that AUC and overall accuracy Pc highest of the invention, and when corresponding same verification and measurement ratio 0.9, this hair Bright false alarm rate be it is minimum, illustrate under complex scene, identification performance of the invention is more preferable than existing method.

Experiment 2:

2.1) experiment scene:

This experiment sectioning image used miniSAR data set disclosed in the U.S. laboratory Sandia, these numbers The website in the laboratory Sandia is downloaded under, partially sliced example images are as shown in figure 4, Fig. 4 (a) is target class training slice map As example, Fig. 4 (b) is clutter class sectioning image example, and Fig. 4 (c) is test sectioning image example.

2.2) experimental selection with experiment 1 identical four groups of traditional characteristics:

2.3) experiment parameter:

Training clutter number of slices p₁=1531, training objective number of slices p₂=2080, test clutter number of slices k₁=510, it surveys Try target slice number k₂=79, weighting parameters λ=0.1, scale factor, α=50, weighting parameters η₀₁=η₁₁=η₂₁=η₀₂ =η₁₂=η₂₂=0.05, dictionary learning the number of iterations iterMax=15, sample weights update interval iterSkip=5, dictionary Atomicity n₀=n₁=n₂=300, weight limit value w_m=50, SVM classifier uses LIBSVM kit, SVM punishment in experiment Coefficient C=10；

2.4) content of the test:

It is identical with experiment 1.

The identification result of experiment 2 is as shown in table 2:

The identification result of 2 distinct methods of table

Distinct methods	AUC	Pc (Thr=0)	Pd (Thr=0)	Pf (Thr=0)	Pd (Thr corresponds to Pd=0.9)	Pf (Thr corresponds to Pd=0.9)
							Verbout	0.8671	75.7216%	0.8734	0.2608	0.8987	0.2980
Verbout+Gao	0.8225	65.0255%	0.8354	0.3784	0.8987	0.5333
							Lincoln	0.8359	67.0628%	0.8861	0.3627	0.8987	0.3784
Lincoln+Verbout+Gao	0.7131	64.5121%	0.7342	0.3686	0.8987	0.5294
							CSDL	0.8757	86.2479%	0.5190	0.0843	0.8987	0.2784
The present invention	0.8923	86.2479%	0.5063	0.0824	0.8987	0.2490

As seen from Table 2, AUC of the invention and overall accuracy Pc highest, and when corresponding same verification and measurement ratio 0.9, this hair Bright false alarm rate be it is minimum, illustrate under complex scene, identification performance of the invention is more preferable than existing method.

Experiment 3:

3.1) experiment scene:

This experiment sectioning image used miniSAR data set disclosed in the U.S. laboratory Sandia, these numbers The website in the laboratory Sandia is downloaded under, partially sliced example images are as shown in figure 5, Fig. 5 (a) is target class training slice map As example, Fig. 5 (b) is clutter class sectioning image example, and Fig. 5 (c) is test sectioning image example.

3.2) experimental selection with experiment 1 identical four groups of traditional characteristics.

3.3) experiment parameter:

Training clutter number of slices p₁=1414, training objective number of slices p₂=1567, test clutter number of slices k₁=627, it surveys Try target slice number k₂=159, weighting parameters λ=0.1, scale factor, α=50, weighting parameters η₀₁=η₁₁=η₂₁=η₀₂ =η₁₂=η₂₂=0.05, dictionary learning the number of iterations iterMax=15, sample weights update interval iterSkip=5, dictionary Atomicity n₀=n₁=n₂=300, weight limit value w_m=50, SVM classifier uses LIBSVM kit, SVM punishment in experiment Coefficient C=10；

3.4) experiment content:

It is identical with experiment 1.

The identification result of experiment 3 is as shown in table 3:

The identification result of 3 distinct methods of table

Distinct methods	AUC	Pc (Thr=0)	Pd (Thr=0)	Pf (Thr=0)	Pd (Thr corresponds to Pd=0.9)	Pf (Thr corresponds to Pd=0.9)
							Verbout	0.5688	42.4936%	0.8428	0.6810	0.8994	0.7927
Verbout+Gao	0.5662	42.4936%	0.8428	0.6810	0.8994	0.7927
							Lincoln	0.5663	44.5293%	0.9623	0.6858	0.8994	0.6284
Lincoln+Verbout+Gao	0.5751	43.1298%	0.9560	0.7018	0.8994	0.6268
							CSDL	0.8529	75.5729%	0.7987	0.2552	0.8994	0.3907
The present invention	0.8555	77.4809%	0.7799	0.2265	0.8994	0.3652

As seen from Table 3, AUC of the invention and overall accuracy Pc highest, and when corresponding same verification and measurement ratio 0.9, this hair Bright false alarm rate be it is minimum, illustrate under complex scene, identification performance of the invention is more preferable than existing method.

Experiment 4:

4.1) experiment scene:

This experiment sectioning image used miniSAR data set disclosed in the U.S. laboratory Sandia, these numbers The website in the laboratory Sandia is downloaded under, partially sliced example images are as shown in fig. 6, Fig. 6 (a) is target class training slice map As example, Fig. 6 (b) is clutter class sectioning image example, and Fig. 6 (c) is test sectioning image example.

4.2) experimental selection with experiment 1 identical four groups of traditional characteristics:

4.3) experiment parameter:

Clutter class trains number of slices p₁=1736, target class training number of slices p₂=2044, clutter class testing number of slices k₁= 305, target class testing number of slices k₂=115, weighting parameters λ=0.1, scale factor, α=50, weighting parameters η₀₁=η₁₁ =η₂₁=η₀₂=η₁₂=η₂₂=0.05, dictionary learning the number of iterations iterMax=15, sample weights update interval iterSkip =5, dictionary atomicity n₀=n₁=n₂=300, weight limit value w_m=50, SVM classifier uses LIBSVM tool in experiment Packet, SVM penalty coefficient C=10；

4.4) experiment content:

It is identical with experiment 1.

The identification result of experiment 4 is as shown in table 4:

The identification result of 4 distinct methods of table

Distinct methods	AUC	Pc (Thr=0)	Pd (Thr=0)	Pf (Thr=0)	Pd (Thr corresponds to Pd=0.9)	Pf (Thr corresponds to Pd=0.9)
							Verbout	0.7508	77.3810%	0.5043	0.1246	0.8957	0.5443
Verbout+Gao	0.7382	76.6667%	0.4957	0.1311	0.8957	0.5836
							Lincoln	0.8922	86.6667%	0.9913	0.1803	0.8957	0.1541
Lincoln+Verbout+Gao	0.8933	84.5238%	0.8957	0.1738	0.8957	0.1738
							CSDL	0.9456	88.8095%	0.8174	0.0852	0.8957	0.1213
The present invention	0.9508	88.8095%	0.8087	0.0820	0.8957	0.1148

As seen from Table 4, AUC of the invention and overall accuracy Pc highest, and when corresponding same verification and measurement ratio 0.9, this hair Bright false alarm rate be it is minimum, illustrate under complex scene, identification performance of the invention is more preferable than existing method.

To sum up, the present invention is that solved multiple based on the specific SAR target discrimination method with shared dictionary of sample weighting classification SAR target under miscellaneous scene identifies problem, and High Resolution SAR image detailed information abundant is effectively utilized, improves complexity SAR target under scene identifies performance.

Claims (9)

1. including: based on the specific SAR target discrimination method with shared dictionary of sample weighting classification

(1) using SAR-SIFT descriptor to given training sectioning imageWith test sectioning imageLocal feature is extracted, the local feature for training sectioning image is obtainedWith the local feature of test sectioning imageWherein,Indicate clutter class training sectioning image,Indicate target class training sectioning image,Indicate clutter Class testing sectioning image,Indicate target class testing sectioning image,It is clutter class training sectioning image Local feature,It is the local feature of target class training sectioning image,It is clutter class testing slice The local feature of image,It is the local feature of target class testing sectioning image, p₁Indicate clutter class training slice map As number, p₂Indicate target class training sectioning image number, k₁Indicate clutter class testing sectioning image number, k₂Indicate target class Test sectioning image number；

(2) by the clutter class training sectioning image local feature in (1) resulting XAs clutter class training sample, Target class trains sectioning image local featureAs target class training sample, Global Dictionary U is obtained；

2a) initialize clutter category dictionary U₁, target category dictionary U₂, shared dictionary U₀, clutter class training sample weightAnd mesh Mark class training sample weightIf current iteration number iter=0；

2b) according to the clutter category dictionary U under current iteration number₁, target category dictionary U₂With shared dictionary U₀, calculate clutter class instruction Practice slice local featureRarefaction representation coefficient H₁With target class training slice local feature's Rarefaction representation coefficient H₂；

2c) according to 2b) obtained H₁And H₂, using alternate optimization method, update clutter category dictionary U₁, target category dictionary U₂With it is shared Dictionary U₀, obtain updated clutter category dictionary U₁', target category dictionary U₂' and shared dictionary U₀′；

Iter=iter+1 2d) is enabled, current the number of iterations is recorded, judges whether to sample weights update, if mod (iter, IterSkip) it is equal to 0, executes step 2e) it is trained sample weights update；Otherwise, it updates, enables without training sample weight U₁=U₁′、U₂=U₂′、U₀=U₀' return step 2b), wherein iterSkip indicates that training sample weight updates interval, and mod is indicated It takes the remainder；

2e) utilize 2c) obtain U₁′、U₂' and U₀' update clutter class training sample weightObtain updated clutter class instruction Practice sample weightsUtilize 2c) obtain U₁′、U₂' and U₀' update target class training sample weightIt is updated Target class training sample weight afterwards

2f) judge whether current iteration number iter is less than maximum number of iterations iterMax, if being less than, enables U₁=U₁′、U₂= U₂′、U₀=U₀′、Return step 2b), if being equal to, iteration stopping is obtained Final Global Dictionary U=[U₀′,U₁′,U₂′]；

(3) the Global Dictionary U for utilizing (2) to obtain obtains training the local feature X and test sectioning image of sectioning image to (1) Local feature Y carry out standardized sparse coding respectively, obtain train sectioning image local feature code coefficientWith test sectioning image local feature code coefficient:

(4) the local feature coding of the local feature code coefficient V for the training sectioning image for obtaining (3) and test sectioning image Coefficient W carries out feature merging and dimensionality reduction respectively, obtained training sectioning image global characteristics:

With the global characteristics of test sectioning image

(5) using global characteristics V " ' one two class Linear SVM classifier of training of training sectioning image, trained point is used Class device classifies to the global characteristics W " ' of test sectioning image, obtains the categorised decision value of each test sectioning image The categorised decision value decision is compared with the threshold value Thr=0 of setting, if decision >=Thr, recognizes by decision It is otherwise clutter class slice to be target class slice.

2. according to the method described in claim 1, wherein step 2a) in initialize clutter category dictionary U₁, target category dictionary U₂, altogether Enjoy dictionary U₀, clutter class training sample weightWith target class training sample weightIt carries out as follows:

2a1) from10000 local features are randomly selected, with K-SVD algorithm to clutter category dictionaryJust Beginningization, with Lagrange duality algorithm by U₁It updates once, wherein d indicates the dimension of training sectioning image local feature, n₁It indicates Clutter category dictionary atom number；

2a2) from10000 local features are randomly selected, with K-SVD algorithm to target category dictionaryJust Beginningization, with Lagrange duality algorithm by U₂It updates once, wherein n₂Indicate target category dictionary atom number；

2a3) fromWith10000 local features are randomly selected, with K-SVD algorithm to shared dictionaryInitialization, with Lagrange duality algorithm by U₀It updates once, wherein n₀Indicate shared dictionary atom number；

2a4) by clutter class training sample weightWith target class training sample weightIt is initialized to 1.

3. according to the method described in claim 1, wherein step 2b) in calculate clutter class training slice local featureRarefaction representation coefficient H₁With target class training slice local featureRarefaction representation coefficient H₂, It carries out as follows；

Following optimization problems 2b1) are solved by feature-sign search algorithm, obtain i-th of clutter class training slice map The local feature of pictureRarefaction representation coefficient

Wherein i=1 ..., p₁, λ expression weighting parameters, | | | |_FIndicate F norm, | | | |₁Indicate l₁Norm,

The local feature of all clutter class training sectioning images is solvedAfter rarefaction representation coefficient, after obtaining update Clutter class training sectioning image local feature rarefaction representation coefficient

Following optimization problems 2b2) are solved by feature-sign search algorithm, obtain j-th of target class training slice map The local feature of pictureRarefaction representation it is sparse

Wherein j=1 ..., p₂,

The local feature of all target class training sectioning images is solvedAfter rarefaction representation coefficient, after obtaining update Target class training sectioning image local feature rarefaction representation coefficient 。

4. according to the method described in claim 1, wherein step 2c) in update clutter category dictionary U₁, carry out as follows；

Following optimization problems 2c1) are solved by alternate optimization method, update clutter class U₁Dictionary obtains updated clutter class word Allusion quotation U₁':

s.t.||U₁(:,b₁)||₂=1, b₁=1 ..., n₁

Wherein, η₁₁WithIt is weighting parameters, | | | |₂It is l₂Norm, | | | |_FIt is F norm, Be size be n₀List Bit matrix,Be size be n₁×n₀Null matrix,Be size be n₁Unit matrix,Be size be n₀×n₁Zero Matrix, n₁It is clutter category dictionary U₁Atomicity, n₂It is target category dictionary U₂Atomicity, n₀It is shared dictionary U₀Atomicity, n =n₀+n₁+n₂, W₁It is clutter class training sample weight matrix:

m₁=n_L×p₁It is the local feature sum of clutter class training sectioning image, n_LIndicate that part is special in a trained sectioning image Levy number.

5. according to the method described in claim 1, wherein step 2c) in update target category dictionary U₂, carry out as follows；

Following optimization problems 2c2) are solved by alternate optimization method, update target category dictionary U₂, obtain updated target class word Allusion quotation U₂':

s.t.U₂(:,b₂)₂=1, b₂=1 ..., n₂

Wherein, η₂₁WithIt is weighting parameters, | | | |₂It is l₂Norm, | | | |_FIt is F norm, Be size be n₀'s Unit matrix,Be size be n₂×n₀Null matrix,Be size be n₂Unit matrix,Be size be n₀×n₂ Null matrix, n₁It is clutter category dictionary U₁Atomicity, n₂It is target category dictionary U₂Atomicity, n₀It is shared dictionary U₀Atom Number, n=n₀+n₁+n₂, W₂It is target class training sample weight matrix:

m₂=n_L×p₂It is the local feature sum of target class training sectioning image, n_LIndicate the part in a trained sectioning image Characteristic Number.

6. according to the method described in claim 1, wherein step 2c) in update shared dictionary U₀, carry out as follows；

Following optimization problems 2c3) are solved by alternate optimization method, update shared dictionary U₀, obtain updated shared dictionary U₀':

s.t.||U₀(:,b₀)||₂,b₀=1 ..., n₀

Wherein, η₀₁WithIt is weighting parameters, | | | |₂It is l₂Norm, | | | |_FIt is F norm,n₁It is Clutter category dictionary U₁Atomicity, n₂It is target category dictionary U₂Atomicity, n₀It is shared dictionary U₀Atomicity, n=n₀+n₁+ n₂, Be size be n₀'s Unit matrix,Be size be n₁×n₀Null matrix,Be size be n₁Unit matrix,Be size be n₀×n₁'s Null matrix, W₁It is clutter class training sample weight matrix:

m₁=n_L×p₁It is the local feature sum of clutter class training sectioning image, n_LIndicate that part is special in a trained sectioning image Number is levied, It is big Small is n₀Unit matrix,Be size be n₂×n₀Null matrix,Be size be n₂Unit matrix,It is that size is n₀×n₂Null matrix, W₂It is target class training sample weight matrix:

m₂=n_L×p₂It is the local feature sum of target class training sectioning image.

7. according to the method described in claim 1, wherein step 2e) in update clutter class training sample weightAnd target Class training sample weightIt carries out as follows:

2e1) utilize 2c) obtain U₁′、U₂' and U₀' update clutter class training sample weightObtain updated clutter class The weight of training sampleWherein i-th of clutter class training sample weight w_1i' obtained by following equations,

In formula, i=1 ..., p₁, α is a scaling factor bigger than 1, w_mIt is the maximum value in weight allowed band, It is the local feature X of i-th of clutter class training sectioning image₁ ⁱRarefaction representation coefficient, value utilize feature-sign Search algorithm solving optimization problemIt obtains,It is U₀' corresponding sparse table Show coefficient,It is U₁' corresponding rarefaction representation coefficient,It is U₂' corresponding rarefaction representation coefficient,It is clutter Sectioning image local featureUse target category dictionary U₂The average energy of ' reconstruct, n_LIndicate office in a trained sectioning image Portion's Characteristic Number；

2e2) utilize 2c) obtain U₁′、U₂' and U₀' update target class training sample weightObtain updated target class The weight of training sampleWherein j-th of target class training sample weight w_2j' obtained by following equations,

In formula, j=1 ..., p₂,It is the local feature X of j-th of target class training sectioning image₂ ^jRarefaction representation coefficient, Value utilizes feature-sign search algorithm solving optimization problem:

It obtains,It is U₀' corresponding rarefaction representation coefficient,It is U₁' corresponding rarefaction representation coefficient,It is U₂' corresponding Rarefaction representation coefficient,It is target slice image local featureUse clutter category dictionary U₁The average energy of ' reconstruct Amount.

8. according to the method described in claim 1, wherein to the local feature code coefficient V of training sectioning image in step (4) Feature merging and dimensionality reduction are carried out, is carried out as follows:

Training sectioning image 4a) is divided into three sons that size is 1 × 1,2 × 2,4 × 4 using spatial pyramid Matching Model Region A1, A2, A3；

4b) merged using maximum value and carries out subregion A1, the local feature code coefficient V of the corresponding training sectioning image of A2, A3 Merge and the global feature of sectioning image trained in splicing, formation:

To the global feature V ' carry out l of training sectioning image₂Norm normalizes, the training sectioning image after being normalized Global feature V ", wherein h indicates the dimension of global characteristics after feature merging；

Dimensionality reduction, the global characteristics of the training sectioning image after obtaining dimensionality reduction 4c) are carried out to V " using principal component analysisWherein h ' is the dimension of the global characteristics after dimensionality reduction.

9. according to the method described in claim 1, wherein to the local feature code coefficient W of test sectioning image in step (4) Feature merging and dimensionality reduction are carried out, is carried out as follows:

4d) will test sectioning image to be divided into size using spatial pyramid Matching Model is 1 × 1,2 × 2,4 × 4 these three sons Region B1, B2, B3；

4e) merged using maximum value by subregion B1, the local feature code coefficient W of the corresponding test sectioning image of B2, B3 into Row merges and splicing, forms the global feature of test sectioning image:

To the global feature W ' carry out l of test sectioning image₂Norm normalizes, the test sectioning image after being normalized Global characteristics W "；

Dimensionality reduction 4f) is carried out to W " using principal component analysis, the global characteristics of the test sectioning image after obtaining dimensionality reductionWherein h ' is the dimension of the global characteristics after dimensionality reduction.