CN103870829A - SAR image-based vehicle target feature extraction method - Google Patents

Abstract

The invention provides an SAR image-based vehicle target feature extraction method. The technical scheme comprises the following steps of Step 1, performing windowing processing on an SAR image to be measured by utilizing the given side length of a smooth window, and calculating the mass Mi,j,L in the smooth window at any position (i,j) in the SAR image to be measured; Step 2, calculating the possibility of the mass Mi,j,L in the smooth window; Step 3, calculating the clearance degree feature of a vehicle target. The clearance degree feature extracted by the method has an affine invariant feature, is strong in identification performance, and is wide in application rage; furthermore, the method only relates to certain simple mathematical calculation, is small in calculation amount, and is convenient for engineering realization.

Description

Vehicle target feature extracting method based on SAR image

Technical field

The invention belongs to sensor information processing technology field, relate to a kind of SAR(synthetic aperture radar, synthetic-aperture radar) extracting method of image vehicle target feature.

Background technology

SAR is a kind of active microwave imaging sensor, has round-the-clock and round-the-clock ability to work, can realize large-area high-resolution imaging by platforms such as spaceborne, airborne or UAV systems and survey, and is widely used in earth observation field.Vehicle target as car, truck, tank, panzer etc. be all a class target of emphasis monitoring in SAR earth observation, the detection of such target and discriminating are the important contents of SAR image information acquisition.

In the detection of SAR image vehicle target and discrimination process, the normal CFAR detection strategy realize target that adopts detects, owing to only having considered the statistical property of background clutter, its testing result may comprise the false target that a large amount of background atural object (as the Sudden change region of background, isolated rock, sparse grove etc.) produces.In order to eliminate above-mentioned false target, need to utilize the real features of vehicle target and priori and the judgement of comparing of the vehicle target feature extracted from SAR image.Therefore, the key of removing false target is just digging vehicle target and the difference of background atural object in SAR image, and quantitatively characterizing, i.e. vehicle target feature extraction.

Conventional SAR image vehicle target feature comprises geometric properties, textural characteristics, contrast metric etc.The extracting method of above-mentioned Partial Feature comes from optical imagery treatment technology, can not truly reflect the real backscattering characteristic of vehicle target in SAR image, and its discriminating performance to background clutter is difficult to guarantee.In addition, the quality of SAR image is easily subject to the impact of the many factors such as platform disturbance, atmospheric scattering, feature changes in actual applications, the SAR image obtaining can produce certain distortion and distortion, can be approximated to be the variations such as the rotation, convergent-divergent, stretching, translation of SAR image, be commonly referred to affined transformation.In the time that distortion and distortion occur SAR image, the result of calculation of above-mentioned most of target signature will have a greater change, and reduce its separable degree to background clutter.Therefore, how to extract the affine invariant features of SAR image vehicle target, and will be that SAR image vehicle target detects and one of key issue of differentiating with this effective differentiation vehicle target and background atural object.

Summary of the invention

The invention provides a kind of vehicle target feature extracting method based on SAR image.The false target that this feature can be distinguished preferably vehicle target and be formed by background atural object, can reduce the false-alarm probability in vehicle target detection and discrimination process largely; And the present invention extract feature there is affine constant characteristic, there is good data adaptability.Meanwhile, this feature calculation is simple, and operand is little, is convenient to Project Realization.

Basic ideas of the present invention are: by analyzing vehicle target and background atural object phenomenology characteristic and the formation mechanism thereof in SAR image, excavate the essential difference that both present on SAR image, the pixel amplitudes that is vehicle target has larger irregular characteristic, and background atural object does not have this characteristic.This irregular characteristic is difficult to utilize euclidean geometry to come quantitative description and calculating, the present invention is based on fractal theory and has proposed a kind of new vehicle target feature with affine invariant feature, namely clearence degree feature.

Technical scheme of the present invention is:

The first step: calculate the quality in smoothing windows

If SAR image to be measured is of a size of W × W, smoothing windows is of a size of L × L, the length of side that L is smoothing windows, L ∈ { 3,5,7,9,11,13}, and L < W.When smoothing windows is above slided in the optional position of SAR image to be measured (i, j), in smoothing windows, pixel maximal value and the minimum value of SAR image to be measured are respectively U _l(i, j) and B _l(i, j), its difference is:

δ _l(i, j)=U _l(i, j)-B _l(i, j) (one)

The mass M of smoothing windows _{i, j, L}be illustrated in SAR picture position (i, j) and locate, in the time that the smoothing windows length of side is L, the fluctuating quantity of image pixel intensities in smoothing windows, is designated as:

M _{i, j, L}=Ceil[k δ _l(i, j)/L] (two)

K=H in formula ₀/ G is weighting coefficient, and G is the pixel maximal value of SAR image to be measured, H ₀be quantization parameter, determine δ _lthe quantification progression of (i, j), gets H conventionally ₀=50, function Ceil[] represent capping round values.

By mass M _{i, j, L}computing formula is known, M _{i, j, L}reflect that position (i, j) locates the degree of irregularity of vehicle target pixel amplitudes in smoothing windows, M _{i, j, L}value is larger, and in smoothing windows, the fluctuating Shaoxing opera of pixel amplitudes is strong, has larger scrambling.

Second step: the probability that calculates quality in smoothing windows

If M ∈ is { M _{i, j, L}, n (M, L) represents to work as mass M _{i, j, L}when=M and the smoothing windows length of side are L, the total number of smoothing windows, mass M _{i, j, L}probability be;

$P(M_{i,j,L}=M)=\frac{n(M,L)}{{(W-L+1)}^2}$ (3)

The 3rd step: the clearence degree feature of calculating vehicle target

Utilize following formula to calculate vehicle target clearence degree feature Λ (L):

$\mathrm{\&Lambda;}\left(L\right)=\frac{\underset{M\&Element;\left\{M_{i,j,L}\right\}}{\mathrm{\&Sigma;}}{\mathrm{<figure-callout\; id="2"\; label="M"\; filenames="HDA0000382920530000031.png"\; state="\{\{state\}\}">M</figure-callout>}}^2\&CenterDot;P(M_{i,j,L}=M)}{{[\underset{M\&Element;\left\{M_{i,j,L}\right\}}{\mathrm{\&Sigma;}}M\&CenterDot;P(M_{i,j,L}=M)]}^2}$ (4)

As a further improvement on the present invention, the vehicle target clearence degree feature of extraction also can be carried out logarithm operation, is designated as:

Λ ' (L)=log[Λ (L)] (five)

Above formula further promotes the performance distinguished and the stability of clearence degree feature.

Further improve as of the present invention, the vehicle target clearence degree feature of extraction also can add up summation on the basis of logarithm operation again, be called clearence degree under polygon elongate member accumulative total with, be designated as:

${\mathrm{\&Lambda;}}^{\&prime;\&prime;}\left(L\right)=\underset{L\&Element;\{\mathrm{3,5,7},...,N_0\}}{\mathrm{\&Sigma;}}{\mathrm{\&Lambda;}}^{\&prime;}\left(L\right)$ (6)

N in above formula ₀≤ 13.Under different length of side L conditions, the clearence degree feature of vehicle target is all greater than the clearence degree feature of background atural object, therefore, accumulates the clearence degree feature under polygon long L condition, can increase the discriminating performance of this feature.

The invention has the beneficial effects as follows:

(1) the irregular characteristic of vehicle target pixel amplitudes in the clearence degree feature energy quantitatively characterizing SAR image that the present invention extracts, can better distinguish vehicle target and background atural object, and then remove the false target by background ground deposits yields, reduce the false-alarm probability that in SAR image, vehicle target detects.

(2) the clearence degree feature that the present invention extracts has been described the distribution situation of M.Point subitem in formula (four)

reflect M _i,jfluctuating quantity, can characterize the scrambling of pixel space tissue and the size of gap size.When trend towards at 0 o'clock, M _{i, j, L}fluctuating less, the spatial group of image pixel is woven with better regularity, the gap size of its regional area is less, conventionally corresponding to natural clutter region.When

when value is larger, M _{i, j, L}there is larger fluctuating, corresponding normally vehicle target region of situation.

(3) checking of the theoretical analysis and measured data, the clearence degree feature that the present invention extracts has affine invariant feature, in the time that testing image generation translation, rotation, convergent-divergent change, the clearence degree eigenwert that this method is calculated remains unchanged, and this affine invariant feature has promoted discriminating performance and the scope of application of clearence degree feature.

(4) in whole implementation process, the present invention only relates to some simple arithmetical operations, and operand is little, is convenient to Project Realization.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of SAR image vehicle target clearence degree feature extracting method provided by the present invention;

Fig. 2 is the SAR image to be measured that comprises vehicle target;

Fig. 3 is the SAR image (not containing vehicle target) of jungle vegetation;

Fig. 4 clearence degree feature logarithm value curve that to be the present invention extract under different smoothing windows length of side conditions Fig. 2 and Fig. 3 data;

Fig. 5 is that the clearence degree feature while utilizing smoothing windows length of side L=3 is differentiated the result of processing to several SAR images to be measured.

Embodiment

Below in conjunction with accompanying drawing, SAR image vehicle target clearence degree feature calculation method provided by the invention is elaborated.

Fig. 1 is the process flow diagram of SAR image vehicle target clearence degree feature calculation method provided by the present invention.The first step of this process flow diagram is to utilize the given smoothing windows length of side to carry out windowing process to SAR image to be measured, calculates the mass M in the smoothing windows that in SAR image to be measured, optional position (i, j) is located _{i, j, L}; Second step is to calculate mass M in smoothing windows according to the formula of second step in technical solution of the present invention (3) _{i, j, L}probability; The 3rd step utilizes the formula (four) of the 3rd step in technical solution of the present invention to calculate the clearence degree feature of vehicle target.

In technical solution of the present invention, the vehicle target clearence degree feature that formula (four) calculates has reflected sliding window mass M _{i, j, L}statistical property.Theoretical analysis shows, in the time that SAR image to be measured does translation, Rotation and Zoom and changes, and mass M in sliding window _{i, j, L}statistical distribution can't change, therefore the present invention calculate clearence degree eigenwert also can not change.That is to say, the clearence degree feature of the SAR image vehicle target that the present invention extracts has translation, Rotation and Zoom invariant feature, has affine invariant feature.

Fig. 2 to Fig. 4 is SAR view data to be measured and the result of utilizing the present invention to test.Fig. 2 is the SAR view data to be measured that comprises a BTR70 vehicle target, and Fig. 3 is the SAR view data to be measured of a jungle atural object, and the resolution of Fig. 2 and Fig. 3 SAR image to be measured is 0.3 meter × 0.3 meter.In the process of calculating, Parameter H ₀=50, sliding window L=3,5 ..., 13, and utilize formula (five) to calculate the logarithm value of clearence degree feature.Fig. 4 clearence degree feature logarithm value that to be the present invention calculate under different smoothing windows length of side conditions Fig. 2 and Fig. 3 data, wherein in figure, adding zero curve is the result of calculation of Fig. 2 data, the curve that adds ▽ is the result of calculation of Fig. 3 data.Obviously, at effectively component-bar chart 2(BTR70 vehicle target of Fig. 4 intermediate gap degree logarithm value) and Fig. 3 (jungle atural object), and have larger separation degree.Certainly, along with the increase of L, both separation degrees reduce, and value all approaches 0, but the clearence degree logarithm value of some vehicle target is all greater than jungle atural object in L≤13.Therefore, we also can utilize the present invention further to improve one's methods, calculate under polygon elongate member clearence degree feature accumulative total and, wherein clearence degree feature accumulative total and be Λ ' ' (L=3 under the polygon elongate member of vehicle target, 5,7 ... 13)=0.225, Λ ' ' (L=3,5,7 of jungle atural object,, 13)=0.144.Under obvious polygon elongate member, clearence degree feature adds up and has larger class spacing, and separability is better than the clearence degree feature of smallest dimension.

Fig. 5 is that the clearence degree feature while utilizing smoothing windows length of side L=3 is differentiated the result of processing to several SAR images to be measured.Experimental data to be measured is to come from MSTAR database.Choosing BMP2, T72 and BTR70 tri-class vehicle targets in MSTAR database is 17 slice of datas while spending at the angle of pitch, has 640, builds the test sample book collection C of vehicle target ₁.In addition, the background clutter data of utilizing MSTAR to provide, can intercept the natural feature on a map slice of data of same size size, build the sample set C of background atural object ₂, have 688.The clearence degree computing method of utilizing the present invention to provide are calculated C in the time of smoothing windows length of side L=3 ₁and C ₂the clearence degree feature of sample set.Result of calculation shows: C ₁the clearence degree feature calculation result of middle vehicle target mainly concentrates on 0.24～0.27; The clearence degree feature of background atural object mainly concentrates on 0.1～0.18.Obviously, to above-mentioned result of calculation, be 0.2 if set clearence degree feature decision threshold, can better distinguish C ₁and C ₂sample.Fig. 5 has provided the discriminating performance statistics of said process intermediate gap degree feature.As shown in Figure 5, clearence degree feature can effectively be distinguished vehicle target and background atural object, it differentiates that performance is better than 98%, there are 2 vehicle targets to be judged to false-alarm, be that false dismissal number of targets is 2, false dismissed rate is about 0.1%, has 8 background atural object sections to be judged to target, be that false target is 8, false dismissed rate is about 0.6%.

Claims (5)

1. the vehicle target feature extracting method based on SAR image, is characterized in that, comprises the steps:

The first step: calculate the quality in smoothing windows:

If SAR image to be measured is of a size of W × W, smoothing windows is of a size of L × L, the length of side that L is smoothing windows, and L < W; When smoothing windows is above slided in the optional position of SAR image to be measured (i, j), in smoothing windows, pixel maximal value and the minimum value of SAR image to be measured are respectively U _l(i, j) and B _l(i, j), its difference is:

δ _l(i, j)=U _l(i, j)-B _l(i, j) (one)

The mass M of smoothing windows _{i, j, L}be illustrated in SAR picture position (i, j) and locate, in the time that the smoothing windows length of side is L, the fluctuating quantity of image pixel intensities in smoothing windows, is designated as:

M _{i, j, L}=Ceil[k δ _l(i, j)/L] (two)

K=H in formula ₀/ G is weighting coefficient, and G is the pixel maximal value of SAR image to be measured, H ₀be quantization parameter, function Ceil[] represent capping round values;

Second step: the probability that calculates quality in smoothing windows:

If M ∈ is { M _{i, j, L}, n (M, L) represents to work as mass M _{i, j, L}when=M and the smoothing windows length of side are L, the total number of smoothing windows, mass M _{i, j, L}probability be;

(3)

The 3rd step: the clearence degree feature of calculating vehicle target:

Utilize following formula to calculate vehicle target clearence degree feature Λ (L):

2. the vehicle target feature extracting method based on SAR image according to claim 1, is characterized in that, extracts vehicle target clearence degree feature Λ ' based on logarithm (L):

Λ′(L)＝log[Λ(L)]。

3. the vehicle target feature extracting method based on SAR image according to claim 2, is characterized in that, extract based on accumulative total and vehicle target clearence degree feature Λ ' ' (L):

N in above formula ₀≤ 13.

4. according to the vehicle target feature extracting method based on SAR image described in claim 1,2, it is characterized in that length of side L ∈ { 3,5,7,9,11, the 13} of smoothing windows.

5. according to the vehicle target feature extracting method based on SAR image described in claim 1,2,3, it is characterized in that, quantization parameter is got H ₀=50.

Images