CN106447688A - Method for effectively segmenting hyperspectral oil-spill image - Google Patents

Abstract

Provided is a method for effectively segmenting a hyperspectral oil-spill image. The method comprises steps of: defining an initial level set function and other related functions; acquiring a new fitting item in combination with a Fisher criterion; constructing an edge stop function to obtain a new length item; performing improvement in combination with an end member extraction algorithm; introducing a level set regular item to prevent reinitialization of the level set function; minimizing an energy function to obtain an Euler-Lagrange equation; setting parameters; selecting a display band and an initial contour; displaying a segmentation result graph; calculating various segmentation precision evaluation indexes; comparing and evaluating the accuracy of the segmentation results. The method can classify a target area in a simulated hyperspectral image and a real hyperspectral image, and effectively segments the hyperspectral oil-spill image with boundary blur and noise, improves the segmentation accuracy of the hyperspectral image, obtain a more accurate classification effect, makes the parameter change more stable, makes the contour curve more accurate, obtains the continuous and closed boundary contour, and has higher precision of segmentation.

Description

A kind of effective dividing method of EO-1 hyperion oil spilling image

Technical field

The present invention relates to a kind of Remote Sensing Image Processing Technology, be specifically related to the dividing method of EO-1 hyperion oil spilling image.

Background technology

The development experience of remote sensing technology panchromatic (black and white), colour phhotograpy, multispectral scanner imaging, and height till now The spectral remote sensing stage.High spectrum resolution remote sensing technique is in brief with narrow and continuous print spectrum channel carries out lasting imaging to atural object Technology, other remote sensing technologies a series of such as it has, and resolution ratio is high, collection of illustrative plates unification, multiband and carrying contains much information are not had New feature.

It is usually used in the segmentation of high spectrum image or sorting algorithm has Threshold segmentation, clustering procedure, region-growing method, support vector Machine method and Active contour.And EO-1 hyperion oil spilling image has the characteristics that：(1) due to the diffusion in oil district, borderline region oil Film thickness is relatively thin, interacts with seawater, causes image Zhong You district obscurity boundary, is therefore not suitable for using dividing based on border Cut algorithm；(2) sea is reflective and the existence of the factor such as fog so that in oil spilling image, profit corresponding curve of spectrum difference is relatively Little, so that EO-1 hyperion oil spilling image segmentation difficulty increases.(3) due to other environmental factors such as equipment self-defect and light Impact, oil spilling image has gray scale inhomogeneities toward contact so that segmentation difficulty increase further.(4) by sunshine and The impact of marine stormy waves so that EO-1 hyperion oil spilling image generally contains speck and shade noise, is therefore also not suitable for using cluster With threshold value etc. based on the dividing method of pixel, and the segmentation that should choose actively profile and region growing etc. based on region is calculated Method.How partitioning algorithm no matter picture quality based on active profile can obtain the border of smooth closing, in continuous oil district There is on partition problem advantage substantially.With the introducing of level set, active contour model develops rapidly especially, various based on Actively the innovatory algorithm of profile emerges in an endless stream so that actively this model of profile is quickly used for the segmentation of all kinds of image.

In conjunction with the active contour model of level set can be divided into again based on border and based on region.Typical based on border Active contour model have geometry active contour model and geodetic active contour model, this class model depends on gradient information, right Noise ratio is more sensitive.Based on the active contour model in region, the non-boundary master that wherein most typically Chan and Vese proposes Dynamic skeleton pattern, i.e. CV model, it has and does not relies on gradient information, partitioning boundary can relatively obscure or contain noisy figure Picture, insensitive to initial curve position, it is easy to accomplish etc. series of advantages, and can enter row vector expand and be applied to height In spectrum picture segmentation.But despite of that, CV model even other any dividing methods are not to have all images There is universality, particularly to the remote sensing oil spilling with features such as obscurity boundary, gray scale inhomogeneities, low contrast, Noises Image is difficult to preferable segmentation effect.Zhang et al. based on the feature of high spectrum image and CV model, propose a kind of based on The region active contour model that spatial spectral limits, referred to as MCV-SSC model (Modified CV model based on Spectral and Space Constrain), and apply on true high spectrum image, it is thus achieved that preferable segmentation effect.

The present invention sets up on CV model, introduces Edge-stopping function and Fisher criterion, establishes a kind of new height Spectrum picture partitioning algorithm, improves segmentation precision while not affecting partitioning algorithm efficiency.

Content of the invention

In order to realize the effective monitoring to marine oil overflow region, the present invention provide a kind of can be in the premise not affecting efficiency Under, improve effective dividing method of the EO-1 hyperion oil spilling image of EO-1 hyperion oil spilling image segmentation precision.

To achieve these goals, technical scheme is as follows：

First, initial level set function and other correlation functions are defined

Initial level set function is commonly defined as：

φ (x, t=0)=± c (1)

Wherein, t=0 represents that initial profile C, x are independent variable, and φ is level set function, if some x is in profile inner function value For-c, otherwise then value is+c.

Definition Heaviside function H and first derivative δ thereof：

Wherein, H represents Heaviside function, and δ represents the first derivative of H function, and z is real number independent variable.Defined function H_ε,δ_εAs follows：

Wherein, z is real number independent variable, and ε is defined as a very little real number, and H represents Heaviside function, and δ represents H letter The first derivative of number, in model energy function, uses function H_ε,δ_εApproximate representation H, δ respectively.

2nd, combine Fisher criterion and obtain new fit term

One good sorting technique of Fisher rule definition should reach between maximum class difference in spacing and minimum class, Can be expressed as with formula：

Wherein,It is respectively the average vector of two class samples,It is respectively the statistical variance of two class samples, when J_fisherOptimal classification effect is reached when taking maximum.

Fit term in former CV model is that variance within clusters is with object function is that variance within clusters is minimum.If by Fisher criterion Basic thought be incorporated into CV model, master mould is replaced variance within clusters, on this basis add inter-class variance) in, then improve After fit term can be expressed as：

Wherein, J_fisherIn conjunction with the fit term function of Fisher principle, (x, y) is pixel coordinate value, and I represents EO-1 hyperion Image,It is respectively the mean vector inside and outside contour curve C, λ₁,λ₂For the constant factor of inside and outside fit term, Ω represents image Region, H_ε,δ_εApproximate representation H respectively, δ, φ represent level set function.

3rd, construct Edge-stopping function and obtain new length item

Spectral modeling refers to the angle between image picture element spectrum vector and sample reference spectra vector, is used for measuring two vectors Between similarity degree, unrelated with the size of two vectors.Its computing formula is as follows：

In formula, SAM is spectral modeling, and N is wave band number.A=(A₁,A₂,…,A_N) and B=(B₁,B₂,…,B_N) difference representative sample The spectral value of two image picture elements in this space, when spectral modeling SAM is less, then show two spectrum vectors closer to.

Edge-stopping function only need to meet a positive function successively decreasing with regard to spectral modeling gradient.So being defined For：

In formula, I represents high spectrum image,For spectral modeling gradient, (x y) represents the coordinate value of pixel, g_sam(α) Spectral modeling gradient function, α represent image any point pixel vector I (x, y) consecutive points pixel vector I (x+ Δ x, y) And I (the spectral modeling distance between x, y+ Δ y), when being in image boundary, the size of spectral modeling gradientValue is relatively big,Value is less.

Introduce the above-mentioned edge function g based on spectral modeling gradient information_sam(α) the length item after, then improving can represent For：

Wherein, L represents length item, and Ω represents image-region, g_sam(α) being spectral modeling gradient function, (x y) represents pixel The coordinate value of point, δ represents the first derivative of Heaviside function, and φ is level set function,Representing the increment of φ, μ is long Degree term coefficient.

4th, the improvement of end member extraction algorithm is combined

CV model is typically only used for two classes and divides, and i.e. can only divide the image into as two regions of target and background.And show Truth condition is often to contain multiclass atural object in image, and when in image containing multiclass atural object, it is right that active contour model cannot realize The division of specific atural object interested.Therefore, CV model is combined with Endmember extraction algorithm, thus realize from containing multiple atural objects Image in mark off specific atural object.

Compared to traditional Endmember extraction algorithm, ATGP algorithm can be in the case of without prior information, from target figure End member vector is extracted in Xiang.ATGP algorithm is a kind of Endmember extraction algorithm theoretical based on non-supervisory Orthogonal subspace projection. It is theoretical that this algorithm is first according to convex geometry, using pixel maximum for brightness in orthogonal subspaces as candidate's end member, thus To initial end member vector, with this constructor space, ask for corresponding orthogonal subspaces, and all pixels are projected to this positive jiao zi Space, asks for projecting the pixel of maximum in orthogonal subspaces as next end member vector；By that analogy, until finding appointment The end member of number.

ATGP algorithm is used to automatically extract target end member.

First extract the end member vector of certain kinds from image with ATGP algorithm, then the position with this end member vector is Central point arranges initial profile, and is replaced in the contour curve improving in active contour model energy functional by this end member vector The mean vector in portion.If setting the certain kinds end member vector that t extracts from high spectrum image as ATGP algorithm, then now improve mould The energy functional of type is：

Wherein, E_tRepresenting objective energy function, α represents the spectral modeling being determined by (7) formula,Represent level set function The increment of φ, (x y) represents the coordinate value of pixel, g_sam(α) it is spectral modeling gradient function, μ, v, λ₁,λ₂Be constant coefficient and Meet μ >=0, v >=0, λ₁,λ₂＞ 0, I be in vector image (x, y) the pixel vector at place, φ is level set function,It is respectively The gray average vector of contour curve C interior zone and perimeter, H_ε,δ_εRepresent approximation Heaviside function and one respectively Order derivative, Ω represents image space.

5th, introducing level set regular terms avoids level set function to reinitialize

In order to avoid level set function reinitializes, saving algrithm time complexity, and then improve efficiency of algorithm, introduce Such as sub-level set regular terms：

Ω represents image space,Represent the increment of level set function φ.

6th, energy functional minimizes and obtains Euler-Lagrange equation

In sum, the energy functional of new established model is

Wherein, E_τRepresenting objective energy function, Ω represents image space, and I is vector image.μ,v,λ₁,λ₂It is constant coefficient And meet μ >=0, v >=0, λ₁,λ₂＞ 0, I (x, y) is vector image, and φ is level set function,Represent level set function The increment of φ.Being respectively the mean vector of contour curve C interior zone and perimeter, t is the end member of corresponding certain kinds Vector,Or t, g_sam(α) it is spectral modeling gradient function, H_ε,δ_εRepresent that approximation Heaviside function and single order thereof are led respectively Number.

Energy functional is minimized and obtains Euler-Lagrange equation, by its mobilism, obtain corresponding gradient descent flow For

Wherein,Or t,Representing the increment of level set function φ, α represents the spectral modeling being determined by (7) formula, μ represents level set regularization coefficient, μ, v, λ₁,λ₂It is constant coefficient and meet μ >=0, v >=0, λ₁,λ₂＞ 0, I are vector image, φ is level set function,Being respectively the mean vector of contour curve C interior zone and perimeter, t is corresponding certain kinds End member vector,Or t, g_sam(α) it is spectral modeling gradient function, δ_εRepresent the first derivative of approximation Heaviside function, Div (x) represents divergence.

7th, parameters is set

Iteration time step delta t, function H_εParameter ε, length term coefficient μ, area term coefficient v, fit term coefficient lambda₁,λ₂, Level set regularization coefficient η.

8th, display wave band and initial profile are selected

The wave band as the final profile of display for the of a relatively high wave band of contrast is selected from several wave bands.Then Initial profile is set, is typically chosen justifying or being uniformly distributed the less multiple circles of radius with the specific center of circle.

9th, segmentation result figure is shown

It by equation discretization, and is iterated by the number of times setting, the final position in the picture of display profile.

Tenth, various segmentation precision evaluation index is calculated

First binaryzation is carried out to the segmentation result image of gained, then in conjunction with real binary image, calculate and make mistakes point Rate, leakage point rate, wrong indexs such as leaking a point rate sum, Kappa coefficient and overall accuracy of dividing.

11, segmentation result precision is compared, evaluates

And for the high spectrum image of simulation high spectrum image and known genuine looks on the spot, we by calculating and can compare Corresponding evaluation index evaluates its segmentation effect；And for the unknown the high spectrum image of true landforms, can only be straight by human eye See and evaluate.

Compared with prior art, the invention have the advantages that：

1st, the present invention is named based on the high-spectrum image segmentation method MCV-FE model (Modified improving CV model CV model based on Fisher Criterion and Edge Function).MCV-FE model is at CV model On the basis of, by introducing Fisher criterion and spectral modeling edge function, and combine end member extraction algorithm, formation a kind of new High-spectrum image segmentation method.Test result indicate that, the method can not only accurately mark off simulation high spectrum image and true height Target area in spectrum, and can effectively split there are obscurity boundary, gray scale inhomogeneities, low contrast, Noise etc. The EO-1 hyperion oil spilling image of feature.Compared with CV model, the present invention can be on the premise of ensureing very not affect efficiency, significantly Improve the segmentation precision of high spectrum image, provide a kind of new method for the accurate oil spilling region that divides.

2nd, in the present invention, MCV-FE model, owing to introducing the basic thought of Fisher criterion, makes the model after improvement not only examine In considering class, error is minimum, considers that between class distance is maximum simultaneously, therefore, it is possible to obtain classifying quality more accurately.In addition, Model after improvement can semi-automatically regulate the ratio of the constant factor of length item and fit term, so that its change to parameter Change more stable.

3rd, CV model is the partitioning algorithm based on region, does not utilize any image edge information, overflows for EO-1 hyperion Oil image, although oils water boundary is very fuzzy, but still has certain effective information.Therefore, consider herein to utilize spectral modeling distance This spectral similarity tolerance one edge indicator function of construction, builds a new length item with this, makes contour curve more Add the border accurately and stably stopping at target area.

4th, the present invention can combine with Endmember extraction algorithm, first obtains the corresponding Seed Points in target area, then evidence This arranges initial profile, and then is capable of the division that the image to region containing multi-class targets carries out certain kinds region.MCV-FE Model, compared with traditional algorithm of region growing and Threshold Segmentation Algorithm, can not only obtain the boundary profile closed continuously, And it is demonstrated experimentally that there is higher segmentation precision.

Brief description

The present invention has 22, accompanying drawing, wherein：

Fig. 1 is simulation EO-1 hyperion oil spilling image the 190th wave band figure (adding noise signal to noise ratio to be 1).

Fig. 2 is the initial profile schematic diagram of analog image.

Fig. 3 is the segmentation result figure for Fig. 1 for the CV model.

Fig. 4 is the segmentation result figure for Fig. 1 for the MCV-FE model.

Fig. 5 is to target with CV model and MCV-FE model Complete Convergence when adding the change of analog image noise signal to noise ratio Change schematic diagram (μ=1, the λ of the iterations of zone boundary₁=λ₂=1).

Fig. 6 is to target with CV model and MCV-FE model Complete Convergence when adding the change of analog image noise signal to noise ratio Change schematic diagram (μ=1, the λ of the iterations of zone boundary₁=λ₂=100).

When Fig. 7 is to change with fit term coefficient lambda value, CV model and MCV-FE model are for Fig. 1 Complete Convergence Change schematic diagram to the iterations of target area boundaries.

Fig. 8 is high spectrum image the 26th wave band figure.

Fig. 9 is high spectrum image atural object mark figure.

Figure 10 is the segmentation result figure for the 1st class target area for the MCV-FE model.

Figure 11 is the segmentation result figure for the 2nd class target area for the MCV-FE model.

Figure 12 is the segmentation result figure for the 3rd class target area for the MCV-FE model.

Figure 13 is the segmentation result figure for the 4th class target area for the MCV-FE model.

Figure 14 is EO-1 hyperion oil spilling image the 190th wave band figure.

Figure 15 is experiment image 1.

Figure 16 is experiment image 2.

Figure 17 is the segmentation result figure corresponding to Figure 15 for the CV model.

Figure 18 is the segmentation result figure corresponding to Figure 15 for the MCV-SSC model.

Figure 19 is the segmentation result figure corresponding to Figure 15 for the MCV-FE model.

Figure 20 is the segmentation result figure corresponding to Figure 16 for the CV model.

Figure 21 is the segmentation result figure corresponding to Figure 16 for the MCV-SSC model.

Figure 22 is the segmentation result figure corresponding to Figure 16 for the MCV-FE model.

Detailed description of the invention

Below in conjunction with the accompanying drawings the present invention is further described.Separately below according to simulation high spectrum image and true height The detailed description of the invention to the present invention for the spectral image data is described.

First, high spectrum image experiment is simulated

In order to verify feasibility and the validity of model, first test on simulation high spectrum image, concrete steps As follows：

A, synthesis EO-1 hyperion simulation oil spilling image

First extracting oil end member and water end (W.E.) unit from true EO-1 hyperion oil spilling image, then synthesizing size is 200 × 200, Wave band number is the simulation high spectrum image of 258, and wherein the part of image middle 100 × 100 is made up of oil end member, other parts By water end (W.E.), unit is constituted, and the noise being eventually adding certain signal to noise ratio (SNR, Signal Noise Ratio) i.e. can obtain used by experiment Analog image.Analog image the 190th band image when Fig. 1 is for adding noise signal to noise ratio to be 2.

B, relative parameters setting

Iteration time step delta t=1, function H_εParameter ε=1, length term coefficient μ=1, area term coefficient v=0, intend Close term coefficient λ₁,λ₂=1 or λ₁,λ₂=1, level set regularization coefficient η=0.2.

C, selection display wave band and initial profile

Select the 190th wave band figure as display image, initial profile is set to be uniformly distributed whole image and a diameter of 5 Multiple roundlets of individual pixel, as shown in Figure 2.

D, display segmentation result figure

When the SNR of the noise that experiment analog image used adds is 2, CV model (iterations is 100) and MCV- The segmentation result figure of FE model (iterations is 30) is as shown in Figure 3, Figure 4：

Fig. 3 is the segmentation result figure for Fig. 1 for the CV model.

Fig. 4 is the segmentation result figure for Fig. 1 for the MCV-FE model.

E, calculate various segmentation precision evaluation index

First, when parameter is set to μ=1, λ₁=λ₂=1, and add analog image noise signal to noise ratio from 0.5 to 10 changes When, iterations when record CV model and MCV-FE model Complete Convergence, as shown in table 1, Fig. 5 is with addition analog image During the change of noise signal to noise ratio, CV model and MCV-FE model Complete Convergence show to the change of the iterations of target area boundaries It is intended to.

Signal to noise ratio snr	0.5	1	2	3	4	5	6	7	8	9	10
												CV model	F	F	37	31	29	27	24	23	23	22	22
MCV-FE model	32	30	19	17	16	16	16	16	15	15	15

Table 1

Then, when parameter is set to μ=1, λ₁=λ₂=100, and add analog image noise signal to noise ratio to become from 0.5 to 10 During change, recording iterations when CV model and MCV-FE model Complete Convergence, as shown in table 2, Fig. 6 is with addition simulation drawing As during the change of noise signal to noise ratio CV model and MCV-FE model Complete Convergence to the change of the iterations of target area boundaries Schematic diagram.

Signal to noise ratio snr	0.5	1	2	3	4	5	6	7	8	9	10
												CV model	36	28	20	18	17	17	17	16	16	16	16
MCV-FE model	32	30	18	17	16	16	15	15	15	15	15

Table 2

Finally, by fit term coefficient lambda₁=λ₂=1 tapers to λ₁=λ₂=100, record CV model and MCV-FE pair Iterations when analog image (SNR=2) Complete Convergence is to target area boundaries, draws corresponding change schematic diagram such as figure Shown in 7.

Lambda	1	10	20	30	40	50	60	70	80	90	100
												CV model	37	22	21	21	21	21	21	20	20	20	20
MCV-FE model	19	18	18	18	18	18	18	18	18	18	18

Table 3

F, segmentation result precision is compared, evaluates

From Fig. 5, it can be seen that CV model is μ=1 when parameter, λ₁=λ₂=1, when signal noise ratio (snr) of image is less than 2, CV model Target area boundaries cannot be converged to.From Fig. 6, it can be seen that when parameter is μ=1, λ₁=λ₂When=100, CV model can be received Hold back target area boundaries, but during except SNR=2, iterations is still many than MCV-FE model, it may thus be appreciated that CV model pair Very sensitive in the change of parameter.

Complex chart 5 and Fig. 6 understands, MCV-FE model Performance comparision under different parameters is stable；Except indivedual special circumstances Outward, MCV-FE model can converge to target area boundaries with less iterations.

From Fig. 7, it is known that find out CV model with the change of fit term coefficient, performance is very unstable；MCV-FE model Change for parameter is relatively stable, and iterations excursion during Complete Convergence is little, and works as fit term coefficient lambda₁=λ₂=1 Change to λ₁=λ₂When=100, convergence rate is constant.Therefore in summary, CV model is more sensitive for parameter, i.e. segmentation knot Fruit and convergence rate are easily subject to the impact of parameter, and MCV-FE model is more stable for the change of parameter.

2nd, true high spectrum image experiment

For verifying effectiveness of the invention further, one group of true high spectrum image experiment will be provided.

A, image sources

The size of experiment image used is 145 × 145, and wave band number is 220.Fig. 8 is the 26 of experiment high spectrum image Wave band figure, Fig. 9 corresponds to the mark figure of Fig. 8.In image in addition to background area, contain 16 class target areas, therefrom select four Class target is tested.In this experiment energy function, the pixel value mean vector of contoured interior will be replaced with all kinds of target area End member vector.Concrete verification step is as follows：

B, relative parameters setting

Iteration time step delta t=1, function H_εParameter ε=1, area term coefficient v=0, level set regularization coefficient η =0.2.Other parameters, the change of the class according to segmentation can be different in an experiment.

C, selection display wave band and initial profile

Select the 26th wave band figure as display image.From experimental image, first extract end member, then initial profile is arranged Circle for position a diameter of 20 pixels for the center of circle with all kinds of interior end members.

D, display segmentation result figure

Figure 10, Figure 11, Figure 12 and Figure 13 are all kinds of MCV-FE model segmentation result figures respectively.Wherein：

Figure 10 is the 1st classification target segmentation result figure.

Figure 11 is the 2nd classification target segmentation result figure.

Figure 12 is the 3rd classification target segmentation result figure.

Figure 13 is the 4th classification target segmentation result figure.

E, calculate various segmentation precision evaluation index

First, thresholding method, region-growing method, MCV-SSC model and the corresponding RMC of MCV-FE model are calculated (by mistake Point rate), RWC (wrong point rate), RWM (sum that mistake misclassification divides) and KC (Kappa coefficient) etc., as shown in table 4.

After wrong point rate RWC (Rate of wrong classification) refers to split, mistake in image target area is divided Pixel number accounts for the percentage of realistic objective area pixel point number.

Leakage in image target area after point rate RMC (Rate of miss classification) refers to segmentation by mistake divides Pixel number accounts for the percentage of realistic objective area pixel point number.

Available mistake point rate and leakage point rate sum RWM (Rate of wrong classification and miss Classification) size weighs the effect of image segmentation, and the value of RWM is less, and the segmentation precision of respective algorithms is described Higher.

Kappa COEFFICIENT K C is a kind of method calculating nicety of grading, and it is by inspection authentic specimen and experiment sample Uniformity realizes, its computing formula is expressed as

KC=(P_o-P_c)/(1-P_c)

P_o=s/n

P_c=(a₁*b₁+ a₀*b₀)/(n*n)

Wherein image total pixel number is n, and realistic objective region pixel number is a₁, background area pixel number is a₀；Sorted Objective area in image pixel number is b₁, background area pixel number is b₀, the equal pixel number of the corresponding pixel value of two width images is s, KC is Kappa coefficient, P_oFor actual concordance rate, P_cFor theoretical concordance rate.

The result of calculation of Kappa COEFFICIENT K C is typically distributed across-1 to 1 in the range of this, but would generally fall between zero and one, Following five kinds of situations can be divided into represent the uniformity of different stage according to value size：Between 0.0-0.20, extremely low one Cause property；Between 0.21-0.40, general uniformity；Between 0.41-0.60, medium uniformity；Between 0.61-0.80, highly Uniformity；Between 0.81-1, almost completely the same.Namely the value of Kappa COEFFICIENT K C is bigger, the nicety of grading of respective algorithms is also Higher.

Overall accuracy OA (Overall Accuracy) refers to that the pixel sum correctly classified accounts for target area sum Percentage.

Then, overall accuracy OA is calculated on this basis again, as shown in table 5.

F, segmentation result precision is compared, evaluates

As can be seen from Table 4, the segmentation precision of thresholding method is much lower relative to other partitioning algorithms, MCV- SSC model takes second place, and MCV-FE model and algorithm of region growing segmentation precision are higher.MCV-FE model is for the 1st class, the 3rd class And the 4th the dividing precision of class atural object the highest, algorithm of region growing takes second place；And for the division of the 2nd class atural object, region growing is calculated Method precision is the highest, and MCV-FE model is compared with algorithm of region growing, and Kappa coefficient is lower slightly.In summary, except at the 2nd class atural object When, Kappa coefficient is slightly below outside algorithm of region growing, and the segmentation precision of improved model is the highest herein.

Finally, overall segmentation accuracy OA calculating each algorithm is as shown in table 5, again, it is seen that MCV-FE mould from table Type overall segmentation accuracy is the highest, and algorithm of region growing takes second place, and third, Threshold Segmentation Algorithm overall segmentation is smart for MCV-SSC model Spend minimum.

The segmentation precision contrast of 4 four kinds of algorithms of table

The overall accuracy of 5 four kinds of methods of table compares

3rd, EO-1 hyperion oil spilling imaging experiments

A, image sources

The true EO-1 hyperion oil spilling image that this experiment is used i.e. derives from 19-3C oil field, Peng Lai, Bohai Sea oil spill accident and is obtained The high-spectral data taking, its size is 512 × 3904, has 258 wave bands.It is this high-spectral data the 190th ripple shown in Figure 14 The part sectional drawing of section image, comprises the classifications such as oil spilling, seawater, ship and drilling platforms in figure, dark parts is oil spilling region, right After it carries out logarithm residual error pretreatment, therefrom intercept experiment image used.Figure 15 is that the part intercepting from Figure 14 contains only excessive Oil and the image of seawater.Figure 16 is the image containing classifications such as oil spilling, seawater, ship and drilling platforms intercepting from Figure 14.This In experiment, when processing Figure 16, in energy functional, the mean vector of contoured interior will be replaced with the end member in oil spilling region Vector.Concrete verification step is as follows：

Concrete verification step is as follows：

B, relative parameters setting

Iteration time step delta t=1, function H_δParameter ε=1, area term coefficient v=0, level set regularization coefficient η =0.2.Other parameters, the change according to trial image in an experiment can be different.

C, selection display wave band and initial profile

Select the 190th wave band figure as display image, initial profile is set to be uniformly distributed whole image and a diameter of 5 Multiple roundlets of individual pixel.

D, display segmentation result figure

Corresponding to Figure 15 3 kinds of models experimental result picture as in figs. 17-19, wherein：

Figure 17 is the segmentation result figure corresponding to Figure 15 for the CV model.

Figure 18 is the segmentation result figure corresponding to Figure 15 for the MCV-SSC model.

Figure 19 is the segmentation result figure corresponding to Figure 15 for the MCV-FE model.

Corresponding to Figure 16 3 kinds of models experimental result picture as illustrated in figs 20-22, wherein：

Figure 20 is the segmentation result figure corresponding to Figure 16 for the CV model.

Figure 21 is the segmentation result figure corresponding to Figure 16 for the HISM model.

Figure 22 is the segmentation result figure corresponding to Figure 16 for the MCV-FE model.

E, calculate various segmentation precision evaluation index

It because the real-time dynamic characteristic of EO-1 hyperion oil spilling image, is difficult to provide sign picture, so cannot with regard to this type of image Calculate its corresponding segmentation precision evaluation index, above-mentioned analysis of image content can only be passed through.

F, segmentation result is compared, evaluate

By Figure 17-19, it can be seen that what CV model was partitioned into is region brighter in the images such as wave, it is impossible to effectively draw Divide oil spilling region；MCV-SSC model segmentation effect is preferable, but is only able to find the part oil spilling border on the left of image, it is impossible to complete Mark off oil spilling region wholely；MCV-FE model in the present invention can be partitioned into oil spilling region exactly.

By Figure 20-22, it can be seen that CV model partition goes out to creep into part brighter in the image such as platform and contamination removing ship, it is impossible to Correctly mark off oil spilling region；MCV-SSC model can relatively accurately find the border on the left of oil spilling region, and cannot find Border on the right side of oil spilling region, and also there is the outline line of many mistakes；MCV-FE model can relatively accurately mark off Oil spilling region, and intra-zone integrality is preferable.

In summary, MCV-FE model can relatively accurately mark off oil spilling region from actual EO-1 hyperion oil spilling image Border, the MCV-FE model in conjunction with end member extraction algorithm also can divide from the actual EO-1 hyperion oil spilling image of atural object containing multiclass Go out oil spilling region interested.

It is above the description carrying out respectively according to simulation, the concrete steps to the present invention for the true hyperspectral image data.This Invention is on the basis of CV model, introduces Fisher criterion and forms new fit term, secondly adds the edge based on spectral modeling Stop function and constitute new length item, improve then in conjunction with Endmember extraction algorithm, finally give a kind of new high-spectrum The partitioning algorithm MCV-FE model of picture.First, in order to prove feasibility and the validity of the present invention, one group of simulation height is carried out Spectrum picture is tested.Understanding from analog image experimental result, the segmentation precision of MCV-FE model is higher, and for the change of parameter Change more stable.

Secondly, in order to verify the validity of MCV-FE model further, the present invention has carried out one group of true high spectrum image Experiment.From Figure 10-13, it can be seen that MCV-FE model can mark off target area interested exactly.From table 1, table 2 can Know, except when the 2nd class atural object, Kappa coefficient is slightly below outside algorithm of region growing, and the segmentation precision of improved model is the highest herein. Thus again demonstrate effectiveness of the invention.

Finally, the present invention applies on Peng Lai 19-3C platform actual EO-1 hyperion oil spilling image.Complete the pretreatment to it it After, intercept parts of images as experimental image.From experimental result picture it can be seen that compared with CV model and HISM model, this MCV-FE model in invention can more accurately mark off oil spilling region.

Simulation and the experiment of actual high spectrum image all show, the MCV-FE model that the present invention proposes can extremely not affect On the premise of segmentation efficiency, can more accurately and stably mark off target area.

Simulation and the experiment of actual high spectrum image all show, the MCV-FE model that the present invention proposes can not affect segmentation On the premise of speed, can more accurately and stably mark off target area.

Claims (1)

1. effective dividing method of an EO-1 hyperion oil spilling image, it is characterised in that:

First, initial level set function and other correlation functions are defined

Initial level set function is commonly defined as：

φ (x, t=0)=± c (1)

Wherein, t=0 represents that initial profile C, x are independent variable, and φ is level set function, if some x profile inner function value for- C, on the contrary then value is+c,

Definition Heaviside function H and first derivative δ thereof：

$\begin{array}{c}H\left(z\right)=\left\{\begin{array}{c}1,z\&GreaterEqual;0\\ 0,z<0\end{array}\right.\\ \mathrm{\&delta;}\left(z\right)=\frac{d}{dz}H\left(z\right)\end{array}---\left(2\right)$

Wherein, H represents Heaviside function, and δ represents the first derivative of H function, and z is real number independent variable, defined function H_ε,δ_εAs Under：

$\begin{array}{c}{H}_{\mathrm{\&epsiv;}}\left(z\right)=\frac{1}{2}(1+\frac{2}{\mathrm{\&pi;}}\arctan (\frac{z}{\mathrm{\&epsiv;}}\left)\right)\\ {\mathrm{\&delta;}}_{\mathrm{\&epsiv;}}\left(z\right)\frac{d}{dz}{H}_{\mathrm{\&epsiv;}}\left(z\right)\end{array}---\left(3\right)$

Wherein, z is real number independent variable, and ε is defined as a very little real number, and H represents Heaviside function, and δ represents H function First derivative, in model energy function, uses function H_ε,δ_εApproximate representation H, δ respectively；

2nd, combine Fisher criterion and obtain new fit term

One good sorting technique of Fisher rule definition should reach between maximum class difference in spacing and minimum class, by public affairs Formula can be expressed as：

$min\left\{\frac{{\mathrm{\&sigma;}}_1^2+{\mathrm{\&sigma;}}_2^2}{{(\stackrel{\&OverBar;}{c_1}-\stackrel{\&OverBar;}{c_2})}^2}\right\}---\left(4\right)$

Wherein,It is respectively the average vector of two class samples,It is respectively the statistical variance of two class samples, work as J_fisherTake Optimal classification effect is reached during maximum；

Fit term in former CV model be variance within clusters and, object function is that variance within clusters is minimum, basic by Fisher criterion Thought is incorporated into CV model, and master mould is replaced variance within clusters, adds inter-class variance on this basis) in, then the plan after improving Close item can be expressed as：

$\begin{array}{c}{J}_{fisher}=\frac{1}{{({\stackrel{\&OverBar;}{c}}_1-{\stackrel{\&OverBar;}{c}}_2)}^2}\&lsqb;{\mathrm{\&lambda;}}_1{\&Integral;}_{\mathrm{\&Omega;}}{|I(x,y)-{\stackrel{\&OverBar;}{c}}_1|}^2{H}_{\mathrm{\&epsiv;}}\left(\mathrm{\&phi;}\right(x,y\left)\right)dxdy\\ +{\mathrm{\&lambda;}}_2{\&Integral;}_{\mathrm{\&Omega;}}{|I(x,y)-{\stackrel{\&OverBar;}{c}}_2|}^2(1-H_{\mathrm{\&epsiv;}}(\mathrm{\&phi;}\left(x,y\right)\left)\right)dxdy\&rsqb;\end{array}---\left(5\right)$

Wherein, J_fisherIn conjunction with the fit term function of Fisher principle, (x, y) is pixel coordinate value, and I represents high spectrum image,It is respectively the mean vector inside and outside contour curve C, λ₁,λ₂For the constant factor of inside and outside fit term, Ω represents image-region, H_ε,δ_εApproximate representation H respectively, δ, φ represent level set function；

3rd, construct Edge-stopping function and obtain new length item

Spectral modeling refers to the angle between image picture element spectrum vector and sample reference spectra vector, is used for measuring between two vectors Similarity degree, unrelated with the size of two vectors, its computing formula is as follows：

$SAM={\cos }^{-1}\frac{A\&CenterDot;B}{\left|A\right|*\left|B\right|}={\cos }^{-1}\frac{\underset{i=1}{\overset{N}{\&Sigma;}}{A}_i*B_i}{\sqrt{\underset{i=1}{\overset{N}{\&Sigma;}}{A}_i^2}*\sqrt{\underset{i=1}{\overset{N}{\&Sigma;}}{B}_i^2}}---\left(6\right)$

In formula, SAM is spectral modeling, and N is wave band number, A=(A₁,A₂,…,A_N) and B=(B₁,B₂,…,B_N) difference representative sample sky The spectral value of two image picture elements between, when spectral modeling SAM is less, then show two spectrum vectors closer to；

Edge-stopping function only need to meet a positive function successively decreasing with regard to spectral modeling gradient, so being defined as：

$\begin{array}{c}\mathrm{\&alpha;}=\frac{1}{2}({\cos }^{-1}\frac{I(x,y)\&CenterDot;I(x+\mathrm{\&Delta;}x,y)}{||I(x,y)||*||I(x+\mathrm{\&Delta;}x,y)||}+{\cos }^{-1}\frac{I(x,y)\&CenterDot;I(x,y+\mathrm{\&Delta;}y)}{||I(x,y)||*||I(x,y+\mathrm{\&Delta;}y)||})\\ g_{sam}\left(\mathrm{\&alpha;}\right)=\frac{1}{1+\left|\&dtri;\mathrm{\&alpha;}\right|}\end{array}---\left(7\right)$

In formula, I represents high spectrum image,For spectral modeling gradient, (x y) represents the coordinate value of pixel, g_sam(α) it is spectrum Angle gradient function, α represent image any point pixel vector I (x, y) consecutive points pixel vector I (x+ Δ x, y) and I (the spectral modeling distance between x, y+ Δ y), when being in image boundary, the size of spectral modeling gradientValue is relatively big,Value is less,

Introduce the above-mentioned edge function g based on spectral modeling gradient information_sam(α) the length item after, then improving can be expressed as：

$L\left(\mathrm{\&phi;}\right)=\mathrm{\&mu;}{\&Integral;}_{\mathrm{\&Omega;}}{g}_{sam}\left(\mathrm{\&alpha;}\right)\mathrm{\&delta;}\left(\mathrm{\&phi;}\right(x,y\left)\right)|\&dtri;\mathrm{\&phi;}(x,y)|dxdy---\left(8\right)$

Wherein, L represents length item, and Ω represents image-region, g_sam(α) being spectral modeling gradient function, (x y) represents pixel Coordinate value, δ represents the first derivative of Heaviside function, and φ is level set function,Representing the increment of φ, μ is length item Coefficient,

4th, the improvement of end member extraction algorithm is combined

CV model is typically only used for two classes and divides, and i.e. can only divide the image into as two regions of target and background, and reality feelings Condition is often to contain multiclass atural object in image, and when in image containing multiclass atural object, active contour model cannot realize to sense emerging Therefore CV model, is combined by the division of the specific atural object of interest with Endmember extraction algorithm, thus realizes from the figure containing multiple atural objects Specific atural object is marked off in Xiang,

Compared to traditional Endmember extraction algorithm, ATGP algorithm can be in the case of without prior information, from target image Extracting end member vector, ATGP algorithm is a kind of Endmember extraction algorithm theoretical based on non-supervisory Orthogonal subspace projection, this calculation It is theoretical that method is first according to convex geometry, using pixel maximum for brightness in orthogonal subspaces as candidate's end member, thus at the beginning of obtaining Top unit vector, with this constructor space, asks for corresponding orthogonal subspaces, and it is empty that all pixels project to this positive jiao zi Between, ask for orthogonal subspaces projecting the pixel of maximum as next end member vector；By that analogy, until finding appointment number End member；

Use ATGP algorithm to automatically extract target end member, first extract from image with ATGP algorithm the end member of certain kinds to Amount, then centered on the position of this end member vector, point arranges initial profile, and improves driving wheel with this end member vector replacement The mean vector within contour curve in wide model energy functional, extracts as ATGP algorithm if setting t from high spectrum image Certain kinds end member vector, then now the energy functional of improved model is：

$\begin{array}{c}{E}_{\mathrm{\&epsiv;}}(t,{\stackrel{\&OverBar;}{c}}_2,\mathrm{\&phi;})=\mathrm{\&mu;}{\&Integral;}_{\mathrm{\&Omega;}}{g}_{sam}\left(\mathrm{\&alpha;}\right){\mathrm{\&delta;}}_{\mathrm{\&epsiv;}}\left(\mathrm{\&phi;}\right(x,y\left)\right)|\&dtri;\mathrm{\&phi;}(x,y)|dxdy\\ +\mathrm{\&nu;}{\&Integral;}_{\mathrm{\&Omega;}}{E}_{\mathrm{\&epsiv;}}\left(\mathrm{\&phi;}\right(x,y\left)\right)dxdy\\ +\frac{1}{{(t-{\stackrel{\&OverBar;}{c}}_2)}^2}({\mathrm{\&lambda;}}_1{\&Integral;}_{\mathrm{\&Omega;}}{\left|I\left(x,y\right)-t\right|}^2{E}_{\mathrm{\&epsiv;}}\left(\mathrm{\&phi;}\left(x,y\right)\right)dxdy\\ +{\mathrm{\&lambda;}}_2{\&Integral;}_{\mathrm{\&Omega;}}{|I(x,y)-{\stackrel{\&OverBar;}{c}}_2|}^2(1-E_{\mathrm{\&epsiv;}}(\mathrm{\&phi;}\left(x,y\right)\left)\right)dxdy)\\ +\mathrm{\&eta;}{\&Integral;}_{\mathrm{\&Omega;}}{\left(\right|\&dtri;\mathrm{\&phi;}|-1)}^{2}dxdy\end{array}---\left(9\right)$

Wherein, E_εRepresenting objective energy function, α represents the spectral modeling being determined by (7) formula,Represent level set function φ's Increment, (x y) represents the coordinate value of pixel, g_sam(α) it is spectral modeling gradient function, μ, ν, λ₁,λ₂It is constant coefficient and meet μ ≥0,ν≥0,λ₁,λ₂＞ 0, I be in vector image (x, y) the pixel vector at place, φ is level set function, t,It is respectively profile The gray average vector of curve C interior zone and perimeter, H_ε,δ_εRepresent that approximation Heaviside function and single order thereof are led respectively Number, Ω represents image space；

5th, introducing level set regular terms avoids level set function to reinitialize

In order to avoid level set function reinitializes, saving algrithm time complexity, and then improve efficiency of algorithm, introduce as follows Level set regular terms：

${\&Integral;}_{\mathrm{\&Omega;}}\frac{1}{2}{\left(\right|\&dtri;\mathrm{\&phi;}|-1)}^{2}dxdy---\left(10\right)$

Ω represents image space,Represent the increment of level set function φ；

6th, energy functional minimizes and obtains Euler-Lagrange equation

In sum, the energy functional of new established model is

$\begin{array}{c}{E}_{\mathrm{\&epsiv;}}(\stackrel{\&OverBar;}{c},{\stackrel{\&OverBar;}{c}}_2,\mathrm{\&phi;})=\mathrm{\&mu;}{\&Integral;}_{\mathrm{\&Omega;}}{g}_{sam}\left(\mathrm{\&alpha;}\right){\mathrm{\&delta;}}_{\mathrm{\&epsiv;}}\left(\mathrm{\&phi;}\right(x,y\left)\right)|\&dtri;\mathrm{\&phi;}(x,y)|dxdy\\ +\mathrm{\&nu;}{\&Integral;}_{\mathrm{\&Omega;}}{E}_{\mathrm{\&epsiv;}}\left(\mathrm{\&phi;}\right(x,y\left)\right)dxdy\\ +\frac{1}{{(\stackrel{\&OverBar;}{c}-{\stackrel{\&OverBar;}{c}}_2)}^2}({\mathrm{\&lambda;}}_1{\&Integral;}_{\mathrm{\&Omega;}}{\left|I\left(x,y\right)-\stackrel{\&OverBar;}{c}\right|}^2{H}_{\mathrm{\&epsiv;}}\left(\mathrm{\&phi;}\left(x,y\right)\right)dxdy\\ +{\mathrm{\&lambda;}}_2{\&Integral;}_{\mathrm{\&Omega;}}{|I(x,y)-{\stackrel{\&OverBar;}{c}}_2|}^2(1-H_{\mathrm{\&epsiv;}}(\mathrm{\&phi;}\left(x,y\right)\left)\right)dxdy)\\ +\mathrm{\&eta;}{\&Integral;}_{\mathrm{\&Omega;}}{\left(\right|\&dtri;\mathrm{\&phi;}|-1)}^{2}dxdy\end{array}---\left(11\right)$

Wherein, E_εRepresenting objective energy function, Ω represents image space, and I is vector image, μ, v, λ₁,λ₂It is constant coefficient and expire Foot μ >=0, v >=0, λ₁,λ₂＞ 0, I (x, y) is vector image, and φ is level set function,Represent level set function φ's Increment,It is respectively the mean vector of contour curve C interior zone and perimeter, the end member vector that t is corresponding certain kinds,Or t, g_sam(α) it is spectral modeling gradient function, H_ε,δ_εRepresent approximation Heaviside function and first derivative thereof respectively；

Minimizing energy functional and obtaining Euler-Lagrange equation, by its mobilism, obtaining corresponding gradient descent flow is

$\frac{\&part;\mathrm{\&phi;}}{\&part;t}={\mathrm{\&delta;}}_{\mathrm{\&epsiv;}}\&lsqb;{\mathrm{\&mu;g}}_{sam}\left(\mathrm{\&alpha;}\right)div\left(\frac{\&dtri;\mathrm{\&phi;}}{\left|\&dtri;\mathrm{\&phi;}\right|}\right)-\mathrm{\&nu;}-\frac{1}{{(\stackrel{\&OverBar;}{c}-{\stackrel{\&OverBar;}{c}}_2)}^2}({\mathrm{\&lambda;}}_1{\left(I-\stackrel{\&OverBar;}{c}\right)}^2-{\mathrm{\&lambda;}}_2{\left(I-{\stackrel{\&OverBar;}{c}}_2\right)}^2)\&rsqb;+\mathrm{\&eta;}({\&dtri;}^2\mathrm{\&phi;}-div(\frac{\&dtri;\mathrm{\&phi;}}{\left|\&dtri;\mathrm{\&phi;}\right|}\left)\right)---\left(12\right)$

Wherein,Or t,Representing the increment of level set function φ, α represents the spectral modeling being determined by (7) formula, and μ represents Level set regularization coefficient, μ, v, λ₁,λ₂It is constant coefficient and meet μ >=0, v >=0, λ₁,λ₂＞ 0, I are vector image, and φ is Level set function,Being respectively the mean vector of contour curve C interior zone and perimeter, t is the end of corresponding certain kinds Unit's vector,Or t, g_sam(α) it is spectral modeling gradient function, δ_εRepresent the first derivative of approximation Heaviside function, div X () represents divergence；

7th, parameters is set

Iteration time step delta t, function H_εParameter ε, length term coefficient μ, area term coefficient v, fit term coefficient lambda₁,λ₂, level Collection regularization coefficient η；

8th, display wave band and initial profile are selected

Select the wave band as the final profile of display for the of a relatively high wave band of contrast from several wave bands, then arrange Initial profile, is typically chosen justifying or being uniformly distributed the less multiple circles of radius with the specific center of circle；

9th, segmentation result figure is shown

It by equation discretization, and is iterated by the number of times setting, the final position in the picture of display profile；

Tenth, various segmentation precision evaluation index is calculated

First binaryzation is carried out to the segmentation result image of gained, then in conjunction with real binary image, calculate make mistakes a point rate, leakage Rate, wrong point of leakage is divided to divide the indexs such as a rate sum, Kappa coefficient and overall accuracy；

11, segmentation result precision is compared, evaluates

And for the high spectrum image of simulation high spectrum image and known genuine looks on the spot, we by calculating and can compare corresponding Evaluation index evaluate its segmentation effect；And for the unknown the high spectrum image of true landforms, intuitively can only be commented by human eye Valency.