CN104240257A - SAR (synthetic aperture radar) image naval ship target identification method based on change detection technology - Google Patents

Abstract

The invention provides an SAR (synthetic aperture radar) image naval ship target identification method based on a change detection technology. According to the technical scheme, the SAR image naval ship target identification method based on the change detection technology comprises the following steps of firstly, establishing change detection amount by using ROI (region of interest) section gray scale data and generating a standard gray scale difference image; secondly, performing threshold segmentation on the difference image and acquiring the number of potential target pixels and the number of real target pixels; and finally, identifying an ROI section by using pixel aggregation degree of targets according to the ratio of the number of the real target pixels to the number of the potential target pixels. By the SAR image naval ship target identification method based on the change detection technology, the separation capacity on naval ship targets in an SAR image and sea clutter false alarm can be effectively improved, and the adaptive capacity on the identification method can also be improved effectively.

Description

Based on the SAR image Ship Target discrimination method of change detection techniques

Technical field

The present invention relates to a kind of SAR (Synthetic Aperture Radar, synthetic-aperture radar) image Ship Target discrimination method, particularly relate to the SAR image target discrimination method based on change detection techniques.

Background technology

In SAR image, the difference of target and background clutter false-alarm scattering properties makes both present different features.Therefore, the target discrimination method that feature based extracts is the most effective in current all discrimination methods, and its application is also extensive.The principle of work that in SAR image, Ship Target is differentiated is ROI (the Region of Interest to suspected target in SAR image, area-of-interest) cut into slices and carry out feature extraction, and select suitable diagnostic characteristics, judgment condition according to setting is differentiated section, what satisfy condition is Ship Target, and ineligible is clutter false-alarm.

The core of existing SAR image Ship Target discrimination method is: (1), on characteristic basis, extracts the diagnostic characteristics that can reflect target and clutter false-alarm essential difference; (2) by Discr., realize differentiating performance.The feature that the discriminating of current Ship Target mainly relies on comprises geometric properties (area, girth, length breadth ratio, moment of inertia etc.), Electromagnetic Scattering (peak value, scattering center etc.), transform characteristics, local invariant feature etc.But, robustness, accuracy and integrality that the factor meeting effect characteristicses such as the multiplicative coherent speckle noise that SAR image-forming mechanism causes, violent scene contrast change extract.For the SAR image of ocean scenes, Ship Target is often subject to the impact of sea clutter, and object and background is difficult to be separated, and cause differentiating mistake, false-alarm cannot effectively be removed.

Summary of the invention

The object of the invention is to differentiate poor-performing, the shortcoming that robustness is inadequate for existing diagnostic characteristics to Ship Target and Sea background clutter, a kind of SAR image Ship Target discrimination method based on change detection techniques is provided, effectively can improves the separating capacity of Ship Target and sea clutter false-alarm and the adaptive faculty of discrimination method in SAR image.

The thinking of technical solution of the present invention is: first utilize ROI section gradation data to build change detection limit, generate standard grayscale differential image; Then, by carrying out Threshold segmentation to differential image, obtain potential target pixel quantity and real goal pixel quantity; Finally, according to the ratio of real goal pixel quantity with potential target pixel quantity, the pixel concentration class of target is utilized to differentiate ROI section.

For solving the problems of the technologies described above, the invention provides a kind of SAR image Ship Target discrimination method based on change detection techniques, technical scheme comprises following process:

The first step: build change detection limit

The ultimate principle building the utilization of change detection limit is: target is usually located at ROI centre of slice, can think clutter background, therefore utilize the average of the pixel of corner location to clutter background to add up in the pixel of section corner location.

Note SAR image ROI section is I, section is square, be of a size of N × N, in ROI section, location of pixels coordinate is that the gray scale of (i, j) is expressed as I (i, j), wherein i, j=1,2,3...N, following formula is utilized to calculate the gray average μ of four square area of section corner location ₁, μ ₂, μ ₃, μ ₄, wherein M=[N/4], [] represents round operation:

$\left\{\begin{array}{c}{\mathrm{\μ}}_1=\frac{1}{M\×M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\underset{j=1}{\overset{M}{\mathrm{\Σ}}}I(i,j)\\ {\mathrm{\μ}}_2=\frac{1}{M\×M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\underset{j=N-M+1}{\overset{N}{\mathrm{\Σ}}}I(i,j)\\ {\mathrm{\μ}}_3=\frac{1}{M\×M}\underset{i=N-M+1}{\overset{N}{\mathrm{\Σ}}}\underset{j=1}{\overset{M}{\mathrm{\Σ}}}I(i,j)\\ {\mathrm{\μ}}_4=\frac{1}{M\×M}\underset{i=N-M+1}{\overset{N}{\mathrm{\Σ}}}\underset{j=N-M+1}{\overset{N}{\mathrm{\Σ}}}I(i,j)\end{array}\right.$ (formula one)

Then background clutter average μ is:

μ=(μ ₁+ μ ₂+ μ ₃+ μ ₄)/4 (formula two)

Following formula is utilized to calculate likelihood ratio change detection limit η (i, j):

$\mathrm{\η}(i,j)=\frac{\mathrm{\μ}}{I(i,j)+1}+\frac{I(i,j)+1}{\mathrm{\μ}}$ (formula three)

Wherein η (i, j), i, j=1,2,3...N.Utilize linear transformation method, by the span linear transformation of all changes detection limit η (i, j) to [0,255], obtain standard grayscale differential image D (i, j), i, j=1,2 ... N.

Second step: adaptive threshold fuzziness

To standard grayscale differential image D (i, j), make p _kfor the probability of D (i, j)=k, i.e. the probability of occurrence of gray level k, k=0,1 ... 255.To any threshold T ∈ [1,254], following formula is utilized to calculate entropy H (T):

$H\left(T\right)=\ln P_T(1-P_T)+\frac{H_T}{P_T}+\frac{H-H_T}{1-P_T}$ (formula four)

Wherein, $P_T=\underset{k=0}{\overset{T}{\mathrm{\Σ}}}{p}_k,H=-\underset{k=0}{\overset{255}{\mathrm{\Σ}}}{p}_k\·\ln p_k,H_T=-\underset{k=0}{\overset{T}{\mathrm{\Σ}}}{p}_k\ln p_k.$

Order i.e. T ₀it is the optimal threshold that object pixel is separated with clutter background pixel.Use threshold value T ₀segmentation is carried out to standard grayscale differential image D (i, j) and obtains bianry image B: namely gray-scale value is greater than threshold value T ₀pixel assignment be 1, otherwise be 0.In bianry image B, set assignment to be potential target pixel as the pixel of 1, add up potential target pixel quantity, be designated as N ₁.

3rd step: object pixel concentration class is analyzed

Region growing is carried out to bianry image B, calculates the number of bianry image B central area connected pixel, be designated as N ₂, calculate object pixel concentration class empirically set decision threshold t, if ρ > is t, then thinks that current ROI exists Ship Target in cutting into slices, otherwise there is not Ship Target.Wherein, carry out region growing to bianry image B, the number calculating bianry image B central area connected pixel adopts existing method to realize, and preferred process is as follows: 3 × 3 neighborhood windows choosing position, bianry image B center are searched for, the pixel that assignment is 1 if cannot search, then N ₂=0, otherwise be that the pixel of 1 is as initial seed point S using assignment ₀; From Seed Points S ₀set out to S ₀8 neighborhoods search for, the pixel that assignment is 1 if exist, is regarded as new Seed Points, with this principle iterative search, until can not find new Seed Points pixel, add up the quantity of all Seed Points now obtained, i.e. the connected pixel number N of bianry image central area B ₂, N ₂reality is Ship Target pixel quantity.

Adopt the present invention can reach following technique effect:

1, the present invention is building in likelihood ratio change detection limit process the thought relating to change and detect, namely background clutter average and whole ROI slice of data is utilized to build change detection limit, generate standard difference view data, be convenient to Ship Target to separate from clutter background.

2, the present invention passes through computed segmentation threshold value adaptively, can obtain the bianry image comprising potential target, and without the need to priori, not only segmentation effect is desirable but also adaptivity is strong.

3, the present invention uses region growing method to binary image statistics Ship Target pixel quantity, and Seed Points search is simple, region growing speed is fast, target identification efficiency is high.

4, the present invention differentiates section according to object pixel concentration class, and method is sane, counting yield is high, differentiate that performance is good.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of SAR image Ship Target discrimination method;

Fig. 2 is the schematic diagram generating standard grayscale differential image;

Fig. 3 is the result after Threshold segmentation;

Fig. 4 is region growing method result;

Fig. 5 is four groups of ROI slice of data identification experiment results.

Embodiment

Below in conjunction with accompanying drawing, embodiments of the present invention are described in detail.

Fig. 1 is the process flow diagram of SAR image Ship Target discrimination method, comprise build change detection limit, adaptive threshold fuzziness, object pixel concentration class analyze three steps.

Fig. 2 to Fig. 5 is the result utilizing the present invention to carry out emulation experiment.

Fig. 2 is the schematic diagram generating standard grayscale differential image.A (), for carrying out the schematic diagram of background clutter mean value computation in SAR image ROI section, the square area that section corner location solid white line comprises is selected calculating gray average μ ₁, μ ₂, μ ₃, μ ₄region, (b) is the standard grayscale differential image that obtains after linear transformation of change detection limit.

Fig. 3 is the result after Threshold segmentation.This figure is the result split standard grayscale differential image Fig. 2 (b), i.e. bianry image B, in figure assignment be 1 pixel be potential target pixel.

Fig. 4 is region growing method result.In figure, Asterisk marks pixel is search for the initial seed point S obtained when carrying out region growing ₀.Region growing obtain pixel quantity be Ship Target pixel quantity.If really have Ship Target in section, then assignment be 1 pixel be highly concentrated in the connected region of centre of slice, therefore can according to region growing obtain pixel the ratio of quantity and potential target pixel quantity differentiate to cut into slices in whether comprise Ship Target.

Fig. 5 is four groups of ROI slice of data identification experiment results.A (), for needing four the ROI sections carrying out target discriminating, the ROI section of two, the left side comprises Ship Target, and the ROI section of two, the right does not comprise Ship Target, only has false-alarm clutter; B bianry image that () obtains for carrying out Threshold segmentation to ROI section in (a); C connected region that () obtains for carrying out region growing to bianry image in (b).Known by calculating, in (a), the object pixel concentration class ρ of four sections is followed successively by from left to right: 0.445,0.493,0.0045,0.0174.Discrimination threshold t ∈ [0.1,0.4] in this example, best value is 0.30.Therefore, by differentiating, in (a), the ROI section of two, the left side comprises Ship Target, and the ROI section of two, the right does not comprise Ship Target.Can be found out by comparison object pixel concentration class, the object pixel concentration class difference that target slice and clutter are cut into slices is obvious, and therefore, identification result is insensitive to choosing of decision threshold t, illustrates that this discrimination method is sane.

Claims (2)

1., based on a SAR image Ship Target discrimination method for change detection techniques, described SAR refers to synthetic-aperture radar, it is characterized in that, comprises the steps:

The first step, builds change detection limit:

Note SAR image ROI section is I, section is square, is of a size of N × N, and in ROI section, location of pixels coordinate is (i, j) gray scale is expressed as I (i, j), wherein i, j=1,2,3...N, ROI refers to area-of-interest, utilizes following formula to calculate the gray average μ of four square area of ROI section corner location ₁, μ ₂, μ ₃, μ ₄, wherein M=[N/4], [] represents round operation:

$\left\{\begin{array}{c}{\mathrm{\μ}}_1=\frac{1}{M\×M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\underset{j=1}{\overset{M}{\mathrm{\Σ}}}I(i,j)\\ {\mathrm{\μ}}_2=\frac{1}{M\×M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\underset{j=N-M+1}{\overset{N}{\mathrm{\Σ}}}I(i,j)\\ {\mathrm{\μ}}_3=\frac{1}{M\×M}\underset{i=N-M+1}{\overset{N}{\mathrm{\Σ}}}\underset{j=1}{\overset{M}{\mathrm{\Σ}}}I(i,j)\\ {\mathrm{\μ}}_4=\frac{1}{M\×M}\underset{i=N-M+1}{\overset{N}{\mathrm{\Σ}}}\underset{j=N-M+1}{\overset{N}{\mathrm{\Σ}}}I(i,j)\end{array}\right.$ (formula one)

Then background clutter average μ is:

μ=(μ ₁+ μ ₂+ μ ₃+ μ ₄)/4 (formula two)

Following formula is utilized to calculate likelihood ratio change detection limit η (i, j):

$\mathrm{\η}(i,j)=\frac{\mathrm{\μ}}{I(i,j)+1}+\frac{I(i,j)+1}{\mathrm{\μ}}$ (formula three)

Wherein η (i, j), i, j=1,2,3...N; Utilize linear transformation method, span linear transformation likelihood ratio being changed detection limit η (i, j), to [0,255], obtains standard grayscale differential image D (i, j), i, j=1,2 ... N;

Second step, adaptive threshold fuzziness:

To standard grayscale differential image D (i, j), make p _kfor the probability of D (i, j)=k, i.e. the probability of occurrence of gray level k, k=0,1 ... 255; To any threshold T ∈ [1,254], following formula is utilized to calculate entropy H (T):

$H\left(T\right)=\ln P_T(1-P_T)+\frac{H_T}{P_T}+\frac{H-H_T}{1-P_T}$ (formula four)

Wherein, $P_T=\underset{k=0}{\overset{T}{\mathrm{\Σ}}}{p}_k,H=-\underset{k=0}{\overset{255}{\mathrm{\Σ}}}{p}_k\·\ln p_k,H_T=-\underset{k=0}{\overset{T}{\mathrm{\Σ}}}{p}_k\ln p_k;$

Make threshold value use threshold value T ₀segmentation is carried out to standard grayscale differential image D (i, j) and obtains bianry image B: namely gray-scale value is greater than threshold value T ₀pixel assignment be 1, otherwise be 0; In bianry image B, set assignment to be potential target pixel as the pixel of 1, add up potential target pixel quantity, be designated as N ₁;

3rd step, object pixel concentration class is analyzed:

Region growing is carried out to bianry image B, calculates the number of bianry image B central area connected pixel, be designated as N ₂, calculate object pixel concentration class according to actual scene setting decision threshold t, if ρ > is t, then thinks that current ROI exists Ship Target in cutting into slices, otherwise there is not Ship Target.

2. the SAR image Ship Target discrimination method based on change detection techniques according to claim 1, it is characterized in that, region growing is carried out to bianry image B, the process calculating the number of bianry image B central area connected pixel is as follows: 3 × 3 neighborhood windows choosing position, bianry image B center are searched for, the pixel that assignment is 1 if cannot search, then N ₂=0, otherwise be that the pixel of 1 is as initial seed point S using assignment ₀; From Seed Points S ₀set out to S ₀8 neighborhoods search for, the pixel that assignment is 1 if exist, is regarded as new Seed Points, with this principle iterative search, until can not find new Seed Points pixel, add up the quantity of all Seed Points now obtained, be the connected pixel number N of bianry image central area B ₂.