CN105335970A - Infrared image segmentation method based on gradient vector improvement model - Google Patents

Abstract

The invention provides an infrared image segmentation method based on a gradient vector improvement model, and relates to the digital image processing field. The infrared image segmentation method based on a gradient vector improvement model can solve the problems that during the infrared image segmentation process, the prior method is sensitive in nose; the image weak edge segmentation is leaked; and parameter selection is short of adaptivity so as to cause unbalance between edge maintenance and enlargement of capture scope. The infrared image segmentation method based on a gradient vector improvement model comprises: introducing guiding filtering and constructing an edge graph according to a kernel function of the guiding filtering; using the edge graph to construct a weighting function and constructing a forcing function; solving an Euler-Lagrange equation of the forcing function to acquire a final external force field; initializing the contour of the curve; bringing the external force field to energy functional, and solving the Euler-Lagrange equation of the energy functional, after discretization, and using the finite difference mode to perform iteration, wherein the curve evolutes to the target edge under the internal and external combined action. The infrared image segmentation method based on a gradient vector improvement model can effectively realize infrared image segmentation, wherein the infrared images contain noise, low contrast and weak edges.

Description

Based on the Infrared Image Segmentation improving gradient vector model

Technical field

The present invention relates to digital image processing field, being specifically related to a kind of Infrared Image Segmentation based on improving gradient vector model.

Background technology

Significant in the analysis being segmented in infrared image of infrared image and target detection.In recent years, lot of domestic and foreign scholar has done a lot of contribution in infrared Image Segmentation, and proposes a lot of method, as thresholding method, edge detection method and region-growing method etc.But because infrared image has the characteristic at strong noise, low contrast and weak edge, these methods can not realize stable correct segmentation.Therefore infrared image is split accurately and effectively and remain a challenging problem, therefore, study a kind of high-quality Infrared Image Segmentation significant.

First active contour is proposed by people such as Kass.Movable contour model, as a kind of effective method, is widely used in Iamge Segmentation.Movable contour model is initialization evolution curve in interesting image regions, give this curve energy function simultaneously, minimization of energy function makes evolution curvilinear motion, until finally approach image object border, thus obtains closing of the frontier, level and smooth image segmentation result.Energy functional is made up of internal energy term and external energy term.Internal energy term represents curve itself, minimizes it and curve can be made short as far as possible and smooth as far as possible.External energy term contains the data message of image, and minimizing it will make its traction evolution curve close to the object that will split, and finally rests on target edges place.For the structure of external force, existing research work comprises, as Balloon Force, and distance power, gradient vector field (GVF).

In these external force, GVF has larger capture ability and converges to depressed area ability and paid close attention to widely.ε=μ ▽ ²v (x, y)-| ▽ f| ²[V (x, y)-▽ f], wherein V (x, y) is gradient vector flow, and f is outline map, and μ is smoothing parameter.There are some shortcomings in GVF model, as other due to the application of outline map, this model more easily receives noise effect.In the process asking for f, common way is f (x, y)=▽ | G _σ(x, y) * I (x, y) |, although noise can be removed to a certain extent, in infrared Image Segmentation, this will increase the negative effect brought by weak edge, and the border of leakage occurs.In order to better segmentation contains the image at weak edge, in structure outline map, JierongCheng proposes dynamic directional gradient vector flow model and JinshanTang proposes the edge that multidirectional gradient vector flow model makes it possible to identify different directions, A.Kovacs constructs new outline map, (for characterizing edges feature).These improve the impact alleviating noise, but when the noise adulterated in edge is comparatively serious, these methods can not realize infrared Image Segmentation accurately.The selection of this model parameter lacks adaptivity in addition, and the noise content in image is depended in the selection of μ.If μ takes unsuitable value, as too large, then evolution curve will pass weak edge and the segmentation result led to errors.In contrast, if μ value is too small, the external force field obtained will retain too much noise.For above problem, the present invention proposes a kind of dividing method based on gradient vector model of improvement, realizes the segmentation of infrared image.

Summary of the invention

The present invention is for solving existing method when carrying out infrared Image Segmentation, there is noise-sensitive, the segmentation of image weak boundary is revealed, Selecting parameter lacks adaptivity, and then the edge that causes keeps and expands the problem such as unbalance between capture range, provides a kind of Infrared Image Segmentation based on improving gradient vector model.

Based on the Infrared Image Segmentation improving gradient vector model, the method is realized by following steps:

Step one, input original image, and construct outline map e (x, y) according to the original image of input;

Step one by one, introduce and guide filtering, obtain and guide the kernel function W of filtering _xy(I);

Described kernel function W _xy(I) expression formula is:

$W_{xy}\left(I\right)=\frac{1}{{\left|w\right|}^2}\underset{k:(x,y)\∈w_k}{\Σ}1+\frac{(I_x-{\mathrm{\μ}}_k)(I_y-{\mathrm{\μ}}_k)}{{\mathrm{\σ}}^2+\mathrm{\λ}},$ In formula, λ is adjustment parameter, μ _kwith be respectively at window w _kthe average of middle original image I (x, y) and variance, | w| is the number of pixel in window;

The kernel function that step one two, employing obtain does convolution algorithm with each pixel of input original image, and obtains corresponding outline map e (x, y) after asking for gradient; The statement formula of described outline map e (x, y) is:

e(x,y)＝▽|W _xy(I(x,y))*I(x,y)|；

Between step 2, outline map e (x, y) normalization 0 to 1 that step one is constructed, structure weighting function g (x, y), h (x, y), is expressed as with following formula:

$h(x,y)=\left\{\begin{array}{c}1,\left|e\right|\&GreaterEqual;\mathrm{\&tau;}\\ -\frac{e^3}{8{\mathrm{\&tau;}}^3}+\frac{5e}{8\mathrm{\&tau;}}+\frac{1}{2},0<\left|e\right|<\mathrm{\&tau;}\\ 0,\left|e\right|=0\end{array}\right.$

g(x,y)＝1-h(x,y)

In formula, τ is positive number;

Step 3, the weighting function obtained according to outline map and the step 2 of step one acquisition, structure forcing function; And obtain external force field;

The expression formula of described forcing function is:

E(V)＝∫∫g(x,y)*|▽V| ²dxdy+h(x,y)*|V-▽e| ²dxdy

The position of step 4, initialization contour curve;

Step 5, bring in energy functional by the external force field obtained in step 3, adopt finite difference iteration, curve develops and arrives object edge under the effect of interior external force, realizes the segmentation of infrared image.

Beneficial effect of the present invention: Infrared Image Segmentation of the present invention, can effectively realize containing noise, the low and infrared Image Segmentation that edge is more weak of contrast.This algorithm has larger catching range, can converge to sunk area, insensitive to curve-initialized position in addition, has noise robustness and effectively alleviates the phenomenon of weak edge leakage in cutting procedure, can meet actual requirement of engineering.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the Infrared Image Segmentation based on improvement gradient vector model of the present invention;

Fig. 2 is that adopting of the present invention is original image based on edge image comparison figure: 2a before and after image comparison before and after filtering in the Infrared Image Segmentation of improvement gradient vector model and filtering, 2b is the gray scale 3D figure that original image is corresponding, 2c is the outline map of original image, 2d is filtered image, 2e is the gray scale 3D figure that filtered image is corresponding, and 2f is the outline map of filtered image.

Fig. 3 adopts the Infrared Image Segmentation segmentation based on improving gradient vector model of the present invention to contain the segmentation result of the infrared image at weak edge: Fig. 3 a is the original infrared image of input, Fig. 3 b is the segmentation result of image, and Fig. 3 c is the external force field of image in weak edge.

In Fig. 4, Fig. 4 a to Fig. 4 c is segmentation result when adopting the Infrared Image Segmentation based on improving gradient vector model of the present invention initial profile curve to be placed on diverse location, and Fig. 4 d is the external force field of Fig. 4 a.

Fig. 5 a in Fig. 5, Fig. 5 b and Fig. 5 c adopt the Infrared Image Segmentation based on improvement gradient vector model of the present invention to containing noisy infrared Image Segmentation result.

Fig. 6 adopts the Infrared Image Segmentation based on improvement gradient vector model of the present invention to the segmentation result of medical image: 6a, 6b, 6c are medical image, and 6d, 6e are the visible images of other types.

Embodiment

Embodiment one, composition graphs 1 to Fig. 6 illustrate present embodiment, based on the Infrared Image Segmentation improving gradient vector model, comprise the following steps:

Step one, mainly realizes asking for of outline map.Introduce in this process and guide filtering to ask for kernel function, the kernel function expression formula of guiding filtering is: $W_{xy}\left(I\right)=\frac{1}{{\left|w\right|}^2}\underset{k:(x,y)\&Element;w_k}{\&Sigma;}1+\frac{(I_x-{\mathrm{\&mu;}}_k)(I_y-{\mathrm{\&mu;}}_k)}{{\mathrm{\&sigma;}}^2+\mathrm{\&lambda;}}$ Wherein λ is adjustment parameter, μ _kwith for at window w _kthe average of middle input original image I and variance, | the number of pixel in w| window.In this algorithm, navigational figure gets input original image.The outline map finally asked for is e (x, y)=▽ | W _xy(I (x, y)) * I (x, y) |.Wherein I (x, y) is input picture.

In the present embodiment, guide filtering to carry out alternative traditional gaussian filtering by reference, avoid the ill-defined drawback adopting traditional gaussian filtering to bring.As shown in figure 2f, the outline map asked in the present invention, while preserving edge information, avoids the impact of noise.When processing the infrared image containing weak edge, being weakened further if this method avoid edge, alleviating the phenomenon of weak edge leakage.As shown in Figure 3 b, due to the introducing of outline map, final curves stop at edge in evolutionary process, and there will not be the phenomenon of over-segmentation.

Step 2, structure weighting function g (x, y), h (x, y), wherein $\begin{array}{c}g(x,y)=\mathrm{\&phi;}\left(e\right)\\ h\left(x,y\right)=1-g\left(x,y\right)\end{array}$ And external force E (V)=∫ ∫ g (x, y) * | ▽ V| ²dxdy+h (x, y) * | V-▽ e| ²dxdy.In external force, Section 1 is level and smooth item.At homogeneous region namely away from edge, g (x, y)=1, h (x, y)=0, at this moment level and smooth item plays a major role, and this effect is level and smooth vector field.At the place that keeps to the side, g (x, y)=0, h (x, y)=1, at this moment the level and smooth item of Section 1 works hardly, avoids edge by smoothly.Section 2 plays a leading role, and makes V (x, y)=▽ e.Be called zone of transition at edge and homogeneous region, g (x, y), h (x, y) value between zero and one in this region, and this can make vector field extend to homogeneous area continuously from fringe region.In the present invention, weighting function directly to newly to try to achieve outline map relevant, avoids the impact of noise.Normalized to by outline map in function [0,1], the final expression formula of weighting function is:

$h(x,y)=\left\{\begin{array}{c}1,\left|e\right|\&GreaterEqual;\mathrm{\&tau;}\\ -\frac{e^3}{8{\mathrm{\&tau;}}^3}+\frac{5e}{8\mathrm{\&tau;}}+\frac{1}{2},0<\left|e\right|<\mathrm{\&tau;}\\ 0,\left|e\right|=0\end{array}\right.$

g(x,y)＝1-h(x,y)

In addition in the present embodiment, the structure of weighting function directly make use of outline map, and to be different from common methods as what utilize in GGVF be the gradient of outline map.The method, while guarantee segmentation effect, decreases calculated amount.Simultaneously owing to make use of newly-built outline map, weighting function avoids the impact of noise, even if there is noise in image, this algorithm still can ensure larger capture range while edge keeps.

Step 3, asks for external force field.Asking for of external force field can be tried to achieve by minimization E (V), wherein E (V)=∫ ∫ g (x, y) * | ▽ V| ²dxdy+h (x, y) * | V-▽ e| ²dxdy.According to the change differential method, the Euler-Lagrange equation of external force field function can be expressed as: g ▽ ²v-h (V-▽ e)=0, after introducing time variable t, is expressed as this partial differential equation of finite difference scheme iteration,

V (x, y, t+ Δ t)=(1-b Δ t) V (x, y, t)+r [V (x+1, y, t)+[V (x-1, y, t)]+[V (x, y+1, t)]+[V (x, y-1, t)]-4 [V (x, y, t)]]+Δ tc is wherein $\left\{\begin{array}{c}b=h\\ c=b\&dtri;e\end{array}\right.,$ In order to ensure the stability of iterative algorithm, finite difference step-length is

Step 4, composition graphs 4, initialization curve location; Wherein, Fig. 4 a to Fig. 4 c is the segmentation result of initial profile curve when being placed on diverse location.

Step 5, according to the external force field V asked for, is brought in the partial differential equation of the energy functional of model: wherein X is contour curve, and α, β are elasticity and the stiffness coefficient of curve X, and s is parameter of curve.The process of asking for of this equation and external force field to ask for process identical, by method of finite difference iteration, make curve develop to object edge under the acting in conjunction of interior external force, realize segmentation.

Composition graphs 5 and Fig. 6 illustrate present embodiment, are the infrared image of the different doping noise of three width to Fig. 5 a, Fig. 5 b in the simulation result composition graphs 5, Fig. 5 of infrared image, Fig. 5 c.

Choosing of experiment parameter: in order to filtering noise goes out and retains the weak edge of image better, selects suitable window parameter and variance adjustment parameter to be very important in filtering.Consider the characteristic of infrared image, select in split-run test | w|=4, λ=0.04.Parameter τ=0.1 in weighting function.

The segmentation result of infrared image is as shown in Fig. 5 a, Fig. 5 b, Fig. 5 c.Result shows when splitting the infrared image containing noise, weak edge, adopts the method for present embodiment can obtain effective segmentation result.

The dividing method of the infrared image described in present embodiment is also not only confined to the segmentation of infrared image.Fig. 6 is the segmentation result to other types image.The present invention can being observed when processing medical image or other visible images by Fig. 6, also can obtain segmentation result accurately.

Claims (3)

1., based on the Infrared Image Segmentation improving gradient vector model, it is characterized in that, the method is realized by following steps:

Step one, input original image, and construct outline map e (x, y) according to the original image of input;

Step one by one, introduce and guide filtering, obtain and guide the kernel function W of filtering _xy(I);

Described kernel function W _xy(I) expression formula is:

in formula, λ is adjustment parameter, μ _kwith be respectively at window w _kthe average of middle original image I (x, y) and variance, | w| is the number of pixel in window;

The kernel function that step one two, employing obtain does convolution algorithm with each pixel of input original image, and obtains corresponding outline map e (x, y) after asking for gradient; The statement formula of described outline map e (x, y) is:

e(x,y)＝▽|W _xy(I(x,y))*I(x,y)|；

Between step 2, outline map e (x, y) normalization 0 to 1 that step one is constructed, structure weighting function g (x, y), h (x, y), is expressed as with following formula:

g(x,y)＝1-h(x,y)

In formula, τ is positive number;

Step 3, the weighting function obtained according to outline map and the step 2 of step one acquisition, structure forcing function; And obtain external force field;

The expression formula of described forcing function is:

E(V)＝∫∫g(x,y) _*|▽V| ²dxdy+h(x,y)*|V-▽e| ²dxdy

The position of step 4, initialization contour curve;

Step 5, bring in energy functional by the external force field obtained in step 3, adopt finite difference iteration, curve develops and arrives object edge under the effect of interior external force, realizes the segmentation of infrared image.

2. the Infrared Image Segmentation based on improving gradient vector model according to claim 1, is characterized in that, to forcing function in step 3: E (V)=∫ ∫ g (x, y) _*| ▽ V| ²dxdy+h (x, y) * | V-▽ e| ²dxdy, adopts and becomes the differential method, ask for the Euler-Lagrange equation of this function, adopt method of finite difference to carry out iteration, finally obtain external force field after discretize.

3. the Infrared Image Segmentation based on improving gradient vector model according to claim 1, it is characterized in that, the partial differential equation of energy functional described in step 5 is:

In formula, X is contour curve, and α, β are elasticity coefficient and the stiffness coefficient of curve X, and s is parameter of curve.