CN105303548A - SAR (Synthetic Aperture Radar) image feature selection method based on hybrid intelligent optimization algorithm - Google Patents

Abstract

The invention discloses a SAR (Synthetic Aperture Radar) image feature selection method based on a hybrid intelligent optimization algorithm. Fractal features are firstly adopted to enhance the SAR image, and based on a segmented image, an image moment-based azimuth estimation method is provided; then, a Zernike moment, a Gabor wavelet coefficient and a gray level co-occurrence matrix are extracted respectively based on images which are not corrected and after correction to form a candidate feature sequence; a hybrid optimization algorithm combining a genetic algorithm and a two valued particle swarm is adopted to realize SAR image feature selection; and finally, an MSTAR database is adopted to prove effectiveness of the provided algorithm. Experimental results show that the feature set after optimization has certain generalization ability, on one hand, SAR target recognition accuracy is improved, and on the other hand, SAR image target recognition time is reduced.

Description

Based on the SAR image feature selection approach of mixing intelligent optimizing algorithm

Technical field

The invention belongs to target identification and mode identification technology, relate to a kind of SAR image feature selection approach based on mixing intelligent optimizing algorithm.

Background technology

Synthetic-aperture radar (SyntheticApertureRadar, SAR) is a kind of microwave imaging sensor, has round-the-clock, round-the-clock, the feature such as multiband, multipolarization, nationalbe widely used in economy and national defense construction, as the system of defense, marine monitoring system, mineral reserve detection etc. of antiballistic missile.

The key factor affecting SAR image automatic target detection (AutomaticTargetRecognition, ATR) comprises feature extraction and feature selecting, and classifier design two aspect.The feature that current SARATR adopts mainly comprises feature, computer vision characteristic sum electromagnetism top grade based on mathematic(al) manipulation.Feature based on mathematic(al) manipulation has wavelet transformation, PCA, ICA etc.Because this category feature has higher object recognition rate usually, generally directly can use or multiple feature merged.Computer vision feature mainly contains texture, attitude angle, shape, fractal dimension, primary edge etc.; Common electromagnetic signature has scattering center, HRR section etc.This two category feature can correspond to the target in imaging scene, because single tagsort effect is poor, general by using multiple Feature Combination to improve discriminating power.For a multiple feature of SAR image, between them, the redundancy issue of existing characteristics collection, feature set can cross the On The Choice of adjustment, in real time notable feature.Feature selecting is the means that the multiple feature of solution is selected simultaneously, and object filters out the most effective Feature Combination in candidate feature.

Summary of the invention

The present invention is directed to the deficiencies in the prior art, provide a kind of SAR image feature selection approach based on mixing intelligent optimizing algorithm.

Concrete steps of the present invention are:

Step (1). pre-service

1.1SAR image enhaucament

Utilize fractal characteristic, by arranging suitable out to out ε _max, MFFK calculating is carried out to pixel each in original SAR image, and generates corresponding MFFK image, realize the targets improvement in SAR image;

$MFFK(x,y)=\underset{\mathrm{\ϵ}=2}{\overset{{\mathrm{\ϵ}}_{\max }}{\Σ}}{\[K(x,y,\mathrm{\ϵ})-\frac{1}{{\mathrm{\ϵ}}_{\max }-1}\underset{\mathrm{\ϵ}=2}{\overset{{\mathrm{\ϵ}}_{\max }}{\Σ}}K(x,y,\mathrm{\ϵ})\]}^2---\left(1\right)$

Wherein MFFK represents at ε _maxrange scale in D tie up the intensity of variation of area K.

1.2SAR Iamge Segmentation

Threshold segmentation is carried out to image after enhancing, obtains the binary image that object and background is separated.

The attitude of 1.3 azimuthal estimations, image is corrected

Split image to SAR, utilize the position angle of the constant moments estimation target of Hu, computing formula is

$\mathrm{\θ}=\frac{{\tan }^{-1}\left(\frac{2(\frac{M_{11}}{M_{00}}-x_c{y}_c)}{(\frac{M_{20}}{M_{00}}-{x_c}^2)-(\frac{M_{02}}{M_{00}}-{y_c}^2)}\right)}{2}---\left(2\right)$

Wherein M ₀₂, M ₂₀, M ₁₁for second moment, M ₀₀for zeroth order square, (x _c, y _c) represent the barycenter of image.Attitude rectification is carried out to original SAR image in position angle according to estimating.

1.4 cuttings, centralization

61 × 61 image-regions centered by image center are read respectively to original SAR image and SAR segmentation image.

Step (2). feature extraction

2.1 extract the rectangular-shaped feature of Zernike

Segmentation image based on original SAR image and rectification SAR image extracts the rectangular-shaped feature of 34 dimension Zernike respectively.

2.2 extract Zernike square amplitude Characteristics

Segmentation image based on original SAR image and rectification SAR image extracts 34 dimension Zernike square amplitude Characteristics respectively.

2.3 extract position angle feature

Image is split to SAR, utilize Hu not bending moment calculate the position angle of target as position angle feature.

2.4 extract gray level co-occurrence matrixes feature

Segmentation image based on original SAR image and rectification SAR image calculates gray level co-occurrence matrixes feature respectively, using energy, entropy, moment of inertia, relevant average and standard deviation as 8 dimension textural characteristics.

2.4 extract Gabor textural characteristics

Gabor filter has very strong space orientation and set direction, based on original SAR image and the segmentation image correcting SAR image, obtains SAR image 16 tie up Local textural feature by one group of Gabor wavelet.

Step (3). feature selecting

3.1 particles and chromosome coding

Coded system adopts binary coding, and 1 represents that this feature is selected, and 0 represents not selected.

3.2 fitness function designs

Fitness function is from the viewpoint of object recognition rate and recognition time two, and the fitness function adopting the method for weight to obtain represents such as formula (3)

Fitness＝a×AC+b×(1-L ₀/L)(3)

Wherein AC represents the discrimination of current subsequence, L ₀represent the Characteristic Number of current subsequence, L is the total number of feature, and the value of weight coefficient a, b is respectively 0.8,0.2.

3.3 mixing intelligent optimizing algorithms

3.3.1 initialization particle

Adopt position and the speed of the N number of particle of random device initialization.

3.3.2 select outstanding particle

Fitness function value according to N number of particle sorts, and is retained by N/2 high for a fitness function value particle as outstanding particle, and N/2 particle is given up in addition.

3.3.3 particle upgrades

First adopt NBPSO algorithm to upgrade position and the speed of N/2 the particle retained, the speed of particle more new formula is

Wherein represent that the position of particle becomes the probability of 1,0 respectively, work as P _ibstor P _gbstwhen equaling 0, increase, reduce; Otherwise, work as P _ibstor P _gbstwhen equaling 1, reduce, increase, in this way, change 1 He of a certain position of particle table 0direction to keep and for the renewal of particle.The location updating formula of particle is

Wherein v ' _ij(t)=sig (v _ij(t)), represent x _ij(t) negate under scale-of-two, r _ijit is the random value between (0,1).

Then carry out GA operation to new particle and obtain N/2 particle, the renewal particle combinations finally obtained twice becomes N number of particle of future generation, this completes the renewal of whole population.

3.3.4 iteration judges

Judge whether to reach maximum iteration time, if meet, stop iteration, otherwise proceed step 3.3.2 and step 3.3.3.

The present invention is based on the image after not correcting and correcting and extract Zernike square, Gabor wavelet coefficient and gray level co-occurrence matrixes formation candidate feature sequence respectively.Have employed a kind of hybrid optimization algorithm in conjunction with genetic algorithm and two-value population and realize SAR image feature selecting.Experimental result shows, the characteristic set after optimization has certain generalization ability, improves the accuracy rate of SAR target identification on the one hand, reduces the time of SAR image target identification on the other hand.

Accompanying drawing explanation

fig. 1for flow process of the present invention figure.

fig. 2 (a) original for target T72 figure.

fig. 2 (b) be the enhancing of target T72 figure.

fig. 2 (c) be the segmentation of target T72 figure.

fig. 2 (d) be the rectification of target T72 figure.

fig. 3for the flow process of mixing intelligent optimizing algorithm figure.

Embodiment

Below in conjunction with accompanying drawingthe invention will be further described.

as Fig. 1shown in, the present invention includes following steps:

Step (1). pre-service

1.1SAR image enhaucament

Fractal model can match with the surface of natural forms or space structure structure well within the scope of some scale, and this regularity of fractal model is expressed and be there is certain otherness with the surface of man-made target or space structure.According to the difference of fractal characteristic, the man-made target extracted in natural background realizes the enhancing to man-made target.Multi-scale Fractal (multi-scalefractalfeaturerelatedwithK, MFFK) is a fractal parameter measure of variation function, and MFFK can be understood as at ε _maxrange scale in D tie up area (K) intensity of variation, utilize MFFK to realize outstanding man-made target and the difference of natural background on fractal characteristic.By arranging suitable out to out ε _max, MFFK calculating is carried out to pixel each in original SAR image, and generates corresponding MFFK image, realize the targets improvement in SAR image, as Fig. 2 (b) shown in;

$MFFK(x,y)=\underset{\mathrm{\ϵ}=2}{\overset{{\mathrm{\ϵ}}_{\max }}{\Σ}}{\[K(x,y,\mathrm{\ϵ})-\frac{1}{{\mathrm{\ϵ}}_{\max }-1}\underset{\mathrm{\ϵ}=2}{\overset{{\mathrm{\ϵ}}_{\max }}{\Σ}}K(x,y,\mathrm{\ϵ})\]}^2---\left(1\right)$

1.2SAR Iamge Segmentation

Carry out Threshold segmentation to image after enhancing, threshold value is 5, obtains the binary image that object and background is separated, as Fig. 2 (c) shown in.

The attitude of 1.3 azimuthal estimations, image is corrected

Split image to SAR, utilize the position angle of the constant moments estimation target of Hu, computing formula is

$\mathrm{\θ}=\frac{{\tan }^{-1}\left(\frac{2(\frac{M_{11}}{M_{00}}-x_c{y}_c)}{(\frac{M_{20}}{M_{00}}-{x_c}^2)-(\frac{M_{02}}{M_{00}}-{y_c}^2)}\right)}{2}---\left(2\right)$

Wherein M ₀₂, M ₂₀, M ₁₁for second moment, M ₀₀for zeroth order square, (x _c, y _c) represent the barycenter of image.Attitude rectification is carried out to original SAR image in position angle according to estimating, fig. 2 (d) image after correcting according to the position angle estimated is shown.

1.4 cuttings, centralization

61 × 61 image-regions centered by image center are read respectively to original SAR image and SAR segmentation image.

Step (2). feature extraction

2.1 extract the rectangular-shaped feature of Zernike

Zernike square is the projection of piece image on one group of Zernike polynomial expression.Teague etc. for base with complex domain Zernike polynomial expression, obtain and have Zernike square [12] that is orthogonal, invariable rotary characteristic.The heavy Zernike square of n rank m of one width discrete picture I may be defined as

$Z_{n,m}=\frac{(n+1)}{\mathrm{\π}{(N-1)}^2}\underset{x=0}{\overset{N-}{\Σ}}\underset{y=0}{\overset{1N-1}{\Σ}}I(x,y)R_{n,m}\left(\mathrm{\ρ}\right)e^{jm\mathrm{\θ}}---\left(3\right)$

Wherein n=0,1,2 ..., 0≤| m|≤n, and n-|m| is even number.(ρ, θ) is the polar coordinates form of expression under unit circle, R _n,mit is radial polynomial.Zernike square has two very important features: (1) is although Zernike square depends on the translation centralization of target, but amplitude has rotational invariance, namely piece image after rotation amplitude do not change, therefore can with the shape facility of the magnitude extraction SAR image of Zernike square; (2) R _n,mmutually orthogonal, can from the amplitude Characteristics of region of interesting extraction Zernike square as target in the shape situation ignoring target.

Segmentation image based on original SAR image and rectification SAR image extracts the rectangular-shaped feature of 34 dimension Zernike respectively.

2.2 extract Zernike square amplitude Characteristics

Segmentation image based on original SAR image and rectification SAR image extracts 34 dimension Zernike square amplitude Characteristics respectively.

2.3 extract position angle feature

Image is split to SAR, utilize Hu not bending moment calculate the position angle of target as position angle feature.

2.4 extract gray level co-occurrence matrixes feature

Gray level co-occurrence matrixes defines by the joint probability density of two position pixels, it not only reflects the distribution character of image brilliance, also reflects in image the position distribution characteristic between the pixel with same brightness or similar brightness, segmentation image based on original SAR image and rectification SAR image calculates gray level co-occurrence matrixes feature respectively, using energy, entropy, moment of inertia, relevant average and standard deviation as 8 dimension textural characteristics.

2.4 extract Gabor textural characteristics

Gabor filter has very strong space orientation and set direction, based on original SAR image and the segmentation image correcting SAR image, obtain SAR image 16 by one group of Gabor wavelet and tie up Local textural feature, Gabor function medium wavelength is set to 1.5, two parameters of Gaussian function are all set to 0.5.

Step (3). feature selecting

3.1 particles and chromosome coding

Coded system adopts binary coding, and 1 represents that this feature is selected, and 0 represents not selected.

3.2 fitness function designs

Fitness function is from the viewpoint of object recognition rate and recognition time two, and the fitness function adopting the method for weight to obtain represents such as formula (4)

Fitness＝a×AC+b×(1-L ₀/L)(4)

Wherein AC represents the discrimination of current subsequence, L ₀represent the Characteristic Number of current subsequence, L is the total number of feature, and the value of weight coefficient a, b is respectively 0.8,0.2.

3.3 carry out SAR image feature selecting based on mixing intelligent optimizing algorithm, see fig. 3.

3.3.1 initialization particle

Adopt position and the speed of the N number of particle of random device initialization.

3.3.2 select outstanding particle

Fitness function value according to N number of particle sorts, and is retained by N/2 high for a fitness function value particle as outstanding particle, and N/2 particle is given up in addition.

3.3.3 particle upgrades

First adopt NBPSO algorithm to upgrade position and the speed of N/2 the particle retained, the speed of particle more new formula is

Wherein represent that the position of particle becomes the probability of 1,0 respectively, work as P _ibstor P _gbstwhen equaling 0, increase, reduce; Otherwise, work as P _ibstor P _gbstwhen equaling 1, reduce, increase, in this way, change 1 He of a certain position of particle table 0direction to keep and for the renewal of particle.The location updating formula of particle is

Wherein v ' _ij(t)=sig (v _ij(t)), represent x _ij(t) negate under scale-of-two, r _ijit is the random value between (0,1).

Then carry out GA operation to new particle and obtain N/2 particle, the renewal particle combinations finally obtained twice becomes N number of particle of future generation, this completes the renewal of whole population.

3.3.4 iteration judges

Judge whether to reach maximum iteration time, if meet, stop iteration, otherwise proceed step 3.3.2 and step 3.3.3.

Claims (1)

1., based on the SAR image feature selection approach of mixing intelligent optimizing algorithm, it is characterized in that the concrete steps of the method are:

Step (1). pre-service

1.1SAR image enhaucament

Utilize fractal characteristic, by arranging suitable out to out ε _max, MFFK calculating is carried out to pixel each in original SAR image, and generates corresponding MFFK image, realize the targets improvement in SAR image;

$MFFK(x,y)=\underset{\mathrm{\ϵ}=2}{\overset{{\mathrm{\ϵ}}_{max}}{\Σ}}{\[K(x,y,\mathrm{\ϵ})-\frac{1}{{\mathrm{\ϵ}}_{\max }-1}\underset{\mathrm{\ϵ}=2}{\overset{{\mathrm{\ϵ}}_{\max }}{\Σ}}K(x,y,\mathrm{\ϵ})\]}^2---\left(1\right)$

Wherein MFFK represents at ε _maxrange scale in D tie up the intensity of variation of area K;

1.2SAR Iamge Segmentation

Threshold segmentation is carried out to image after enhancing, obtains the binary image that object and background is separated;

The attitude of 1.3 azimuthal estimations, image is corrected

Split image to SAR, utilize the position angle of the constant moments estimation target of Hu, computing formula is

$\mathrm{\θ}=\frac{{\tan }^{-1}\left(\frac{2(\frac{M_{11}}{M_{00}}-x_c{y}_c)}{(\frac{M_{20}}{M_{00}}-{x_c}^2)-(\frac{M_{02}}{M_{00}}-{y_c}^2)}\right)}{2}---\left(2\right)$

Wherein M ₀₂, M ₂₀, M ₁₁for second moment, M ₀₀for zeroth order square, (x _c, y _c) represent the barycenter of image; Attitude rectification is carried out to original SAR image in position angle according to estimating;

1.4 cuttings, centralization

61 × 61 image-regions centered by image center are read respectively to original SAR image and SAR segmentation image;

Step (2). feature extraction

2.1 extract the rectangular-shaped feature of Zernike

Segmentation image based on original SAR image and rectification SAR image extracts the rectangular-shaped feature of 34 dimension Zernike respectively;

2.2 extract Zernike square amplitude Characteristics

Segmentation image based on original SAR image and rectification SAR image extracts 34 dimension Zernike square amplitude Characteristics respectively;

2.3 extract position angle feature

Image is split to SAR, utilize Hu not bending moment calculate the position angle of target as position angle feature;

2.4 extract gray level co-occurrence matrixes feature

Segmentation image based on original SAR image and rectification SAR image calculates gray level co-occurrence matrixes feature respectively, using energy, entropy, moment of inertia, relevant average and standard deviation as 8 dimension textural characteristics;

2.4 extract Gabor textural characteristics

Gabor filter has very strong space orientation and set direction, based on original SAR image and the segmentation image correcting SAR image, obtains SAR image 16 tie up Local textural feature by one group of Gabor wavelet;

Step (3). feature selecting

3.1 particles and chromosome coding

Coded system adopts binary coding, and 1 represents that this feature is selected, and 0 represents not selected;

3.2 fitness function designs

Fitness function is from the viewpoint of object recognition rate and recognition time two, and the fitness function adopting the method for weight to obtain represents such as formula (3)

Fitness＝a×AC+b×(1-L ₀/L)(3)

Wherein AC represents the discrimination of current subsequence, L ₀represent the Characteristic Number of current subsequence, L is the total number of feature, and the value of weight coefficient a, b is respectively 0.8,0.2;

3.3 mixing intelligent optimizing algorithms

3.3.1 initialization particle

Adopt position and the speed of the N number of particle of random device initialization;

3.3.2 select outstanding particle

Fitness function value according to N number of particle sorts, and is retained by N/2 high for a fitness function value particle as outstanding particle, and N/2 particle is given up in addition;

3.3.3 particle upgrades

First adopt NBPSO algorithm to upgrade position and the speed of N/2 the particle retained, the speed of particle more new formula is

Wherein represent that the position of particle becomes the probability of 1,0 respectively, work as P _ibstor P _gbstwhen equaling 0, increase, reduce; Otherwise, work as P _ibstor P _gbstwhen equaling 1, reduce, increase, in this way, the change 1 of a certain position of particle and the direction of table 0 can be kept and for the renewal of particle; The location updating formula of particle is

Wherein v ' _ij(t)=sig (v _ij(t)), represent x _ij(t) negate under scale-of-two, r _ijit is the random value between (0,1);

Then carry out GA operation to new particle and obtain N/2 particle, the renewal particle combinations finally obtained twice becomes N number of particle of future generation, this completes the renewal of whole population;

3.3.4 iteration judges

Judge whether to reach maximum iteration time, if meet, stop iteration, otherwise proceed step 3.3.2 and step 3.3.3.