CN104732551A

CN104732551A - Level set image segmentation method based on superpixel and graph-cup optimizing

Info

Publication number: CN104732551A
Application number: CN201510164483.3A
Authority: CN
Inventors: 王斌; 牛丽军; 关钦; 高新波; 牛振兴; 丁海刚; 吕鑫; 宗汝
Original assignee: Xidian University
Current assignee: Xidian University
Priority date: 2015-04-08
Filing date: 2015-04-08
Publication date: 2015-06-24

Abstract

The present invention relates to a level set image segmentation method based on superpixels and graph cut optimization, which mainly solves the problems in the prior art that the initial position of the level set is sensitive and the segmentation efficiency is low. The implementation steps are as follows: 1. divide the image to be segmented into I block to obtain the superpixel representation of image I; 2. Select a pixel in each superpixel, and use these pixels to form a new image I _s ; 3. According to the Chan-Vese model, construct the level set of the new image I _s Energy functional; 4. Discretize the constructed level set energy functional; 5. Use graph cut technology to optimize the energy functional to realize the segmentation of the new image I _s ; 6. According to the segmentation result of the new image I _s , Complete the segmentation of image I. The invention accelerates the image segmentation speed, reduces the sensitivity to the initialization position of the level set, can obtain the globally optimal segmentation result, and can be used for fast segmentation and recognition of natural images and medical image objects.

Description

The level set image segmentation method of optimization is cut based on super-pixel and figure

Technical field

The invention belongs to technical field of image processing, further relate to a kind of image level collection dividing method, can be used for natural image, the Fast Segmentation of medical image target and identification.

Background technology

In mankind's activity, image is the most frequently used information carrier, and its reason is because the intuitive of image on the one hand, things represents by objectively, without the need to being subject to the impact of description person's subjective factor, being that the mankind can analyze image rapidly and understand on the other hand, obtaining image information instantaneously.

Image Engineering can be divided into image procossing, graphical analysis and image interpretation three part, and image procossing mainly gathers image, obtains image information, and carries out pretreated process to it, the technology such as main application image collection, image enhaucament, filtering; Graphical analysis mainly obtains characteristics of image and describes and interested target, and this process mainly comprises Iamge Segmentation and feature extraction; Image understanding is according to characteristics of image, studies the further feature of image, and this process relates to the contents such as target identification, scene description.Image procossing occupies critical role in Image Engineering, and on the one hand, Iamge Segmentation is the basis of target statement and pattern measurement, plays an important role to graphical analysis; On the other hand, image Segmentation Technology and image is converted into more abstract form based on the objective expression of cutting techniques, characteristic measuring techniques, and greatly reducing the data volume of the required process of image understanding, this makes further high layer analysis, image understanding and artificial intelligence become possibility.Image Segmentation Technology has been widely used in the various aspects of real life, as medical image analysis, remote Sensing Image Analysis, intelligent traffic administration system etc.

Image Segmentation Technology based on level set has become the focus direction splitting area research in recent years, comparatively unified model and framework are formed at present, it allows the topological structure of curve in evolutionary process to change, and also can obtain good segmentation result to multiple goal with the segmentation being communicated with target more.Existing level-set segmentation model has following several method:

1. based on the level-set segmentation methods of image boundary.This method is the velocity function of the gradient information tectonic level collection utilizing image, makes evolution curve approach objective contour.People [the Li C. such as such as Li, Xu C., Gui C.and Fox M., " Levelset evolution without re-initialization:a new variational formulation ", IEEE Conference onComputer Vision and Pattern Recognition, San Diego, CA, USA, pp.430-436, Jun.2005] proposing one need not initialized variational method, and the method, by definition penalty term, makes level set function close to symbolic measurement.The curve evolvement of this level-set segmentation methods based on image boundary depends on image gradient, and the Grad bounded of real image, boundary stops function can not be zero, and curve may pass border, and when picture noise is excessive, segmentation result is also undesirable.

2. based on the level-set segmentation methods of image-region.This method is the energy functional utilizing the area information of image to build level set.People [the T.Chan and L.Vese such as such as Chan, " Active contours without edges ", IEEETransactions on Image Processing, vol.10, no.2, pp.266-277,2001] by Mumford-Shah solution to model is reduced to piecewise constant, propose Chan-Vese model, can be used for the meaningless or ill-defined image of segmentation gradient, but undesirable for the image segmentation that gray scale is uneven.

3. based on the level-set segmentation methods of picture shape priori.This method is the energy functional utilizing the shape prior of image to build level set.People [the Cremers D. such as such as Cremers, Osher S.J.and Soatto S., " Kernel densityestimation and intrinsic alignment for shape priors in level set segmentation ", InternationalJournal of Computer Vision, vol.69, no.3, pp.335-351, Sep.2006] utilize Density Estimator, use shape prior constructs the difference between level set function and given embedding shape, shape difference subitem is joined in Chan-Vese parted pattern, this method does not rely on the restricted hypothesis of Gaussian distribution, allow multiplephase when obvious training shapes is encoded accurately.But this dividing method makes calculated amount greatly increase, reduce the segmentation efficiency of image.

Although the above-mentioned level-set segmentation methods based on image boundary, image-region and shape prior can complete Iamge Segmentation preferably, but these methods are responsive to the initialized location of level set, cannot obtain the segmentation result of global optimum, and the solving speed of model is slow, segmentation efficiency is low.

Summary of the invention

The object of the invention is to the deficiency for above-mentioned prior art, propose a kind of level set image segmentation method cutting optimization based on super-pixel and figure, to improve segmentation efficiency and the segmentation precision of image.

Technical thought of the present invention is: by adopting the super-pixel generation technique of simple linear iteration cluster, accelerate the splitting speed of image, cut the method for optimization by employing figure, reduce the susceptibility of level set being carried out to initialized location, obtain the segmentation result of global optimum.Implementation step comprises as follows:

(1) image I to be split is divided into M block, each block is a super-pixel, obtains the super-pixel SP={sp of image I ₁..., sp _k..., sp _m, wherein sp _krepresent a kth super-pixel, k=1,2 ..., M;

(2) each selection pixel in each super-pixel, utilizes this M pixel to form new images I _s;

(3) according to Chan-Vese model, new images I is built _senergy functional E _cV, it is expressed as

\begin{matrix} E_{CV} = μ {&Integral;}_{Ω} δ (φ) | &dtri; φ | dxdy \\ + λ_{1} {&Integral;}_{Ω} {| I_{s} - c_{in} |}^{2} H (φ) dxdy \\ + λ_{2} {&Integral;}_{Ω} {| I_{s} - c_{out} |}^{2} (1 - H (φ)) dxdy \end{matrix}

Wherein, Ω represents image area, and μ is non-negative parameter, λ ₁and λ ₂be positive parameter, φ represents evolution curve,

H (φ) = \frac{1}{2} + \frac{1}{π} \arctan (\frac{φ}{ϵ}), δ (φ) = \frac{1}{π} \frac{1}{1 + {(φ / ϵ)}^{2}},

ε is constant and ε → 0, c _inand c _outrepresent the inside and outside gray average of closed curve respectively;

(4) to the energy functional E built _cVcarry out discrete statement, its discrete form is as follows:

\begin{matrix} E_{CV}^{d} = μ \underset{p, q}{Σ} w_{pq} (x_{p} (1 - x_{q}) + x_{q} (1 - x_{p})) \\ + λ_{1} \underset{p}{Σ} {| I_{s} - c_{in} |}^{2} x_{p} \\ + λ_{2} \underset{p}{Σ} {| I_{s} - c_{out} |}^{2} (1 - x_{p}) \end{matrix}

Wherein, x _pbe defined as image I _sin the two-valued variable of each pixel, p=(x, y) ∈ Ω, φ (p) represent the value of the level set function at pixel p place, x _pand x _qbe respectively new images I _sin the two-valued variable of two different pixels point p and q, be connected if p with q is axle, weight w _pq=1, be connected if p with q is diagonal angle, then weight represent

(5) use figure cuts technical optimization energy functional, realizes new images I _ssegmentation:

(5.1) the evolution curve of initialization level set;

(5.2) according to the discrete form of level set energy functional the data item E of calculating chart ^p(x _p) and smooth item E ^p,q(x _p, x _q), it is expressed as:

E ^p(x _p)＝λ ₁|I _s-c _in| ²x _p+λ ₂|I _s-c _out| ²(1-x _p)

E ^p,q(x _p,x _q)＝μw _pq(x _p(1-x _q)+x _q(1-x _p))

(5.3) the data item E of figure is utilized ^p(x _p), smooth item E ^p,q(x _p, x _q) and weight matrix design of graphics, obtained the minimal cut of figure by min-cut algorithm, upgrade x _pvalue;

(5.4) compare the energy size that min-cut algorithm uses front and back, if the energy before using the energy after min-cut algorithm to be less than algorithm use, then return step (5.2), continue iterative computation; If both are equal, then stop iteration;

(5.5) according to x _pvalue determination new images I _sthe attribute of middle pixel: if x _p=0, then x _pcorresponding pixel is background, if x _p=1, then x _pcorresponding pixel is target, completes new images I _ssegmentation;

(6) according to new images I _ssegmentation result, complete the segmentation to image I;

New images I _sin each pixel correspondence image I in a super-pixel, if new images I _sin pixel be background, then corresponding in image I super-pixel is also background, if new images I _sin pixel be target, then corresponding in image I super-pixel is also target, thus completes the segmentation to image I.

Compared to the prior art, the present invention has the following advantages:

1) the present invention adopts simple linear iteration clustering technique, generate the super-pixel of image to be split, a pixel is selected in each super-pixel, these pixels selected are utilized to form new image, because the size of new images is far smaller than original image, can calculated amount be greatly reduced in follow-up calculating, improve the efficiency of Iamge Segmentation.

2) traditional energy functional optimization generally adopts gradient descent method, the step-length of the method is selected to have a great impact iterations, and can only local minimum be obtained, the figure that the present invention adopts cuts the mode of technology as a kind of optimization energy functional, the global minimum of energy functional can be tried to achieve, reduce the susceptibility of level set being carried out to initialized location, improve the precision of Iamge Segmentation.

Accompanying drawing explanation

Fig. 1 is realization flow figure of the present invention;

Fig. 2 is the Comparative result figure split natural image with the present invention and Chan-Vese model;

Fig. 3 is the Comparative result figure split noise image with the present invention and Chan-Vese model;

Fig. 4 is the Comparative result figure split artificial image with the present invention and Chan-Vese model.

Embodiment

Below in conjunction with accompanying drawing, 1 couple of the present invention is described in further detail.

Step 1, is divided into M block by image I to be split, and each block is a super-pixel, obtains the super-pixel SP of image I.

The method of existing generation super-pixel has canonical to cut algorithm, mean shift algorithm, rapid drift algorithm and simple linear Iterative Clustering etc., because simple linear Iterative Clustering uses simple, counting yield is high, has developed into the common method that super-pixel generates.The present invention adopts simple linear Iterative Clustering, and generate the super-pixel SP of image I to be split, its step is as follows:

(1.1) rgb color space of image to be split is transformed to CIELAB color space;

(1.2) the expectation number M of super-pixel is inputted;

(1.3) with size be grid interval pixel is sampled, obtain initial cluster centre C _k=[l _k, a _k, b _k, x _k, y _k] ^t, k=1,2 ..., M, wherein, N represents image pixel number, [l _k, a _k, b _k] for CIELAB color space pixel color vector, l _krepresent brightness, a _k, b _krepresent color opposition dimension, x _k, y _kfor location of pixels;

(1.4) according to the minimal gradient position of 3 × 3 neighborhoods, cluster centre is moved on to seed position;

(1.5) initialized pixel i and cluster centre C _kdistance d (i)=∞, label l (i)=-1 of initialization recording pixel i classification;

(1.6) for each cluster centre C _k, calculate C _kin neighbouring 2S × 2S region, each pixel is to cluster centre C _kdistance D, if D < d (i), then make d (i)=D, l (i)=k;

(1.7) calculate new cluster centre and residual error E, wherein residual error E is defined as the L2 norm of successively twice cluster centre position;

(1.8) compare the threshold value T of residual error E and setting, if E≤T, then stop upgrading, otherwise be back to step (1.6), continue to upgrade cluster centre and residual error, until residual error E≤T;

(1.9) pixel being k by the category label that step (1.6) obtains is combined as one piece of super-pixel sp _k, from 1 to M, k is traveled through, is combined into M super-pixel, completes the division to image I, obtain the super-pixel SP={sp of image I ₁..., sp _k..., sp _m, wherein sp _krepresent a kth super-pixel, k=1,2 ..., M.

Step 2, each selection pixel in each super-pixel, utilizes this M pixel to form new images I _s.

(2.1) gray average of each super-pixel is calculated:

μ (k) = \frac{1}{m_{k}} Σ_{j = 1}^{m_{k}} {sp}_{k} (j),

Wherein, μ (k) is a kth super-pixel sp _kgray average, m _krepresent super-pixel sp _kmiddle number of pixels;

(2.2) at super-pixel sp _kin choose | sp _k(j)-μ (k) | the pixel that minimum value is corresponding, k=1,2 ..., M;

(2.3) new images I is formed with the pixel chosen _s.

Step 3, according to Chan-Vese model, builds new images I _slevel set energy functional.

On the basis of Mumford-Shah model, Chan and Vese proposes the level set Image Segmentation Model based on region, the energy functional E of this model _cVbe defined as smooth item E _smoothwith data item E _datasum, that is:

E _CV＝E _smooth+E _data。

For solving energy functional E _cV, introduce Heaviside function and dirac measure wherein, ε is constant and ε → 0, and φ represents evolution curve.

Solve new images I _sthe concrete steps of level set energy functional as follows:

(3.1) the smooth item E of calculated level collection energy functional _smooth:

E _smoothbe defined as the length of closed curve, its computing formula is:

E_{smooth} = μ {&Integral;}_{Ω} δ (φ) | &dtri; φ | dxdy,

Wherein, Ω represents image area, and μ is non-negative parameter, and φ represents evolution curve, for the divergence of φ;

(3.2) the data item E of calculated level collection energy functional _data:

E _data＝λ ₁∫ _Ω|I _s-c _in| ²H(φ)dxdy

，

+λ ₂∫ _Ω|I _s-c _out| ²(1-H(φ))dxdy

Wherein, λ ₁and λ ₂be respectively data item conitnuous forms E _datathe coefficient of Section 1 and Section 2, I _srepresent new images, Ω represents image area, c _inand c _outrepresent the interior gray average of closed curve and outer gray average respectively, account form is as follows:

c_{in} = \frac{{&Integral;}_{Ω} I_{s} H (φ) dxdy}{{&Integral;}_{Ω} H (φ) dxdy}, c_{out} = \frac{{&Integral;}_{Ω} I_{s} (1 - H (φ)) dxdy}{{&Integral;}_{Ω} (1 - H (φ)) dxdy},

(3.3) new images I is calculated _senergy functional E _cV, it is expressed as:

\begin{matrix} E_{CV} = μ {&Integral;}_{Ω} δ (φ) | &dtri; φ | dxdy \\ + λ_{1} {&Integral;}_{Ω} {| I_{s} - c_{in} |}^{2} H (φ) dxdy \\ + λ_{2} {&Integral;}_{Ω} {| I_{s} - c_{out} |}^{2} (1 - H (φ)) dxdy \end{matrix} .

Step 4, to the level set energy functional E built _cVcarry out discrete statement.

According to new images I _slevel set energy functional E _cV, respectively to E _cVsmooth item and data item carry out discrete statement, and concrete steps are as follows:

(4.1) x is defined _pfor new images I _sin the two-valued variable of each pixel, wherein, p=(x, y) ∈ Ω, Ω represents image area;

(4.2) discrete form of the smooth item of level set energy functional is solved

E_{smooth}^{d} = μ \underset{p, q}{Σ} (x_{p} (1 - x_{q}) + x_{q} (1 - x_{p})) w_{pq},

Wherein, μ is non-negative parameter, x _pand x _qbe respectively new images I _sin the two-valued variable of two different pixels point p and q, be connected if p with q is axle, then weight w _pq=1, be connected if p with q is diagonal angle, then weight

(4.3) discrete form of the data item of level set energy functional is solved

E_{data}^{d} = λ_{1}^{'} \underset{p}{Σ} {| I_{s} - c_{in} |}^{2} x_{p} + λ_{2}^{'} \underset{p}{Σ} {| I_{s} - c_{out} |}^{2} (1 - x_{p}),

Wherein, λ ' ₁with λ ' ₂be respectively data item discrete form the coefficient of Section 1 and Section 2, I _srepresent new images, c _inand c _outbe respectively the interior gray average of closed curve and outer gray average, account form is as follows:

c_{in} = \frac{Σ_{p} I_{s} x_{p}}{Σ_{p} x_{p}}, c_{out} = \frac{Σ_{p} I_{s} (1 - x_{p})}{Σ_{p} (1 - x_{p})};

(4.4) level of aggregation collection energy functional E _cVthe discrete form of smooth item with the discrete form of data item obtain E _cVdiscrete form:

\begin{matrix} E_{CV}^{d} = E_{smooth}^{d} + E_{data}^{d} \\ = μ \underset{p, q}{Σ} w_{pq} (x_{p} (1 - x_{q}) + x_{q} (1 - x_{p})) \\ + λ_{1}^{'} \underset{p}{Σ} {| I_{s} - c_{in} |}^{2} x_{p} \\ + λ_{2}^{'} \underset{p}{Σ} {| I_{s} - c_{out} |}^{2} (1 - x_{p}) \end{matrix} .

Step 5, use figure cuts technical optimization energy functional, realizes new images I _ssegmentation.

(5.1) the evolution curve of initialization level set;

E ^p(x _p)＝λ ₃|I _s-c _in| ²x _p+λ ₄|I _s-c _out| ²(1-x _p)，

E ^p,q(x _p,x _q)＝μw _pq(x _p(1-x _q)+x _q(1-x _p))，

Wherein, λ ₃and λ ₄be respectively the data item E of figure ^p(x _p) Section 1 and the coefficient of Section 2;

(5.5) according to x _pvalue determination new images I _sthe attribute of middle pixel: if x _p=0, then x _pcorresponding pixel is background, if x _p=1, then x _pcorresponding pixel is target, completes new images I _ssegmentation.

Step 6, according to new images I _ssegmentation result, complete the segmentation to image I to be split.

New images I _sin the corresponding image I to be split of each pixel in a super-pixel, if new images I _sin pixel be background, then corresponding in image I to be split super-pixel is also background, if new images I _sin pixel be target, then corresponding in image I to be split super-pixel is also target, thus completes the segmentation to image I to be split.

Effect of the present invention can further illustrate by using following emulation experiment

1, simulated conditions

The present invention is Intel (R) Core (TM) i5 2.80GHZ, internal memory 4G, WINDOWS 7 in operating system at central processing unit, uses the emulation that MATLAB software carries out.

2, content is emulated

Emulation 1, with the present invention and existing Chan-Vese model, natural image is split, result as shown in Figure 2, wherein:

Fig. 2 (a) is with Chan-Vese model to the initialization of natural image,

The result that Fig. 2 (b) is split natural image for using Chan-Vese model,

Fig. 2 (c) uses Chan-Vese model to the binary map of the result that natural image is split,

Fig. 2 (d) uses the present invention to the initialization of natural image,

The result that Fig. 2 (e) is split natural image for using the present invention,

Fig. 2 (f) is for using the present invention to the binary map of the result that natural image is split.

Comparison diagram 2 (c) and Fig. 2 (f), can find out that the present invention is good to natural image segmentation effect, can obtain the segmentation result of global optimum.

Emulation 2, with the present invention and existing Chan-Vese model, noise image is split, result as shown in Figure 3, wherein:

Fig. 3 (a) uses Chan-Vese model to the initialization of noise image,

The result that Fig. 3 (b) is split noise image for using Chan-Vese model,

Fig. 3 (c) uses Chan-Vese model to the binary map of the result that noise image is split,

Fig. 3 (d) uses the present invention to the initialization of noise image,

The result that Fig. 3 (e) is split noise image for using the present invention,

Fig. 3 (f) is for using the present invention to the binary map of the result that noise image is split.

Comparison diagram 3 (c) and Fig. 3 (f), can find out that the present invention is to noise robustness, also can obtain good segmentation result under noise existent condition.

Emulation 3, with the present invention and existing Chan-Vese model, artificial image is split, result as shown in Figure 4, wherein:

Fig. 4 (a) uses Chan-Vese model to the initialization of artificial image,

The result that Fig. 4 (b) is split artificial image for using Chan-Vese model,

Fig. 4 (c) uses Chan-Vese model to the binary map of the result that artificial image is split,

Fig. 4 (d) uses the present invention to the initialization of artificial image,

The result that Fig. 4 (e) is split artificial image for using the present invention,

Fig. 4 (f) is for using the present invention to the binary map of the result that artificial image is split.

Comparison diagram 4 (c) and Fig. 4 (f), can find out that the present invention is good to artificial image segmentation effect, can obtain the segmentation result of global optimum.

Emulation 1, emulation 2 and emulation 3 time used and iterations as shown in table 1:

Table 1.

As can be seen from Table 1, compared with Chan-Vese model, the present invention only needs iteration several times just can obtain segmentation result, and the Iamge Segmentation time used is far smaller than Chan-Vese model, and the segmentation efficiency of image has greatly improved.

Claims

1. A level set image segmentation method optimized based on superpixels and graph cuts, comprising the steps of:

(1) Divide the image I to be segmented into M blocks, each of which is a superpixel, and obtain the superpixel SP of the image I = {sp ₁ ,...,sp _k ,...,sp _M }, where sp _k represents the kth Superpixels, k=1,2,...,M;

(2) Select one pixel in each superpixel, and use these M pixels to form a new image I _s ;

(3) According to the Chan-Vese model, the energy functional function E _CV of the new image I _s is constructed, which is expressed as

\begin{matrix} {E E.}_{CV cv} = = {λ λ}_{11} {&Integral; &Integral;}_{Ω Ω} {| | {I I}_{s the s} - - {c c}_{in in} | |}^{22} H h ((φ φ)) dxdy dxdy \\ + + {λ λ}_{22} {&Integral; &Integral;}_{Ω Ω} {| | {I I}_{s the s} - - {c c}_{out out} | |}^{22} ((11 - - H h ((φ φ)))) \\ + + μ μ {&Integral; &Integral;}_{Ω Ω} δ δ ((φ φ)) | | &dtri; &dtri; φ φ | | dxdy dxdy \end{matrix}, dxdy dxdy

Among them, Ω represents the image domain, μ is a non-negative parameter, λ ₁ and λ ₂ are the coefficients of the first and second terms of E _CV respectively, φ represents the evolution curve,

h (φ) = \frac{1}{2} + \frac{1}{π} \arctan (\frac{φ}{ϵ}), δ (φ) = \frac{1}{π} \frac{1}{1 + {(φ / ϵ)}^{2}},

ε is a constant and ε→0, c _in and c _out respectively represent the mean value of the inner gray level and the mean value of the outer gray level of the closed curve;

(4) Discretely express the constructed energy functional E _CV , and its discrete form is as follows:

\begin{matrix} {E E.}_{CV cv}^{d d} = = {λ λ}_{11}^{' '} \underset{p p}{Σ Σ} {| | {I I}_{s the s} - - {c c}_{in in} | |}^{22} {x x}_{p p} \\ + + {λ λ}_{22}^{' '} \underset{p p}{Σ Σ} {| | {I I}_{s the s} - - {c c}_{out out} | |}^{22} ((11 - - {x x}_{p p})) \\ + + μ μ \underset{p p,, q q}{Σ Σ} {w w}_{pq pq} (({x x}_{p p} ((11 - - {x x}_{q q})) + + {x x}_{q q} ((11 - - {x x}_{p p})))) \end{matrix}

Among them, λ′ ₁ and λ′ ₂ are the coefficients of the first and second terms of the discrete form E _CV respectively, and x _p is defined as the binary variable of each pixel in the image I _s , p=(x,y)∈Ω, φ(p) represents the value of the level set function at the pixel point p, x _p and x _q are binary variables of two different pixel points p and q in the new image I _s respectively , if p and q are axially connected, the weight w _pq =1, if p and q are diagonally connected, then the weight

(5) Use the graph cut technology to optimize the energy functional to realize the segmentation of the new image I _s :

(5.1) Initialize the evolution curve of the level set;

(5.2) According to the discrete form of level set energy functional The data item E ^p (x _p ) and the smooth item E ^p,q (x _p ,x _q ) of the calculation graph are expressed as:

E ^p (x _p )＝λ ₃ |I _s -c _in | ² _xp +λ ₄ |I _s -c _out | ² (1-x _p ),

E ^p,q (x _p ,x _q )=μw _pq (x _p (1-x _q )+x _q (1-x _p )),

Wherein, λ ₃ and λ ₄ are the coefficients of the first item and the second item of the data item E ^p (x _p ) of figure respectively;

(5.3) Use the data item E ^p (x _p ) of the graph, the smoothing item E ^p,q (x _p ,x _q ) and the weight matrix to construct the graph, obtain the minimum cut of the graph through the min-cut algorithm, and update the value of x _p ;

(5.4) Compare the energy size before and after the use of the min-cut algorithm, if the energy after using the min-cut algorithm is less than the energy before the use of the algorithm, then return to step (5.2) and continue the iterative calculation; if the two are equal, then stop the iteration;

(5.5) Determine the attributes of the pixels in the new image I _s according to the value of x _p : if x _p =0, then the pixel corresponding to x _p is the background; if x _p =1, then the pixel corresponding to x _p is the target , complete the segmentation of the new image I _s ;

(6) According to the segmentation result of the new image I _s , complete the segmentation of the image I to be segmented.

Each pixel in the new image I _s corresponds to a superpixel in the image to be segmented I, if the pixel in the new image I _s is the background, then the corresponding superpixel in the image I to be segmented is also the background, if the new image If the pixel in I _s is the target, then the corresponding superpixel in the image I to be segmented is also the target, so that the segmentation of the image I to be segmented is completed.

2. the level set image segmentation method based on superpixel and graph cut optimization according to claim 1, wherein in said step (1), the image I to be segmented is divided into M blocks, obtain the superpixel SP of image I, press Follow the steps below:

(1.1) the RGB color space of image to be divided is transformed into CIELAB color space;

(1.2) Input the expected number M of superpixels;

(1.3) Take the size as Pixels are sampled at grid intervals to obtain the initial cluster center C _k =[l _k ,a _k ,b _k ,x _k ,y _k ] ^T , k=1,2,…,M, where N represents The number of image pixels, [l _k , a _k , b _k ] is the pixel color vector of CIELAB color space, l _k represents brightness, a _k , b _k represent color opposite dimensions, x _k , y _k are pixel positions;

(1.4) Move the cluster center to the seed position according to the minimum gradient position of the 3×3 neighborhood;

(1.5) Initialize the distance d(i)=∞ between the pixel i and the clustering center C _k , initialize the label l(i)=-1 of the category of the recorded pixel i, i=1,2,...,N;

(1.6) For each cluster center C _k , calculate the distance D from each pixel in the 2S×2S area near C _k to the cluster center C _k , if D<d(i), then let d(i)= D,l(i)=k;

(1.7) Calculate the new cluster center and residual E, where the residual E is defined as the L2 norm of the two successive cluster center positions;

(1.8) Compare the residual E with the set threshold T, if E≤T, stop updating, if E>T, return to step (1.6), and continue to update the cluster center and residual until the residual E≤ up to T;

(1.9) Combining the pixels with category label k obtained in step (1.6) into a piece of superpixel sp _k , traversing k from 1 to M, combining into M blocks of superpixels, completing the division of image I, and obtaining the image Superpixels SP of I = {sp ₁ , . . . , sp _k , . . . , sp _M }.

3. the level set image segmentation method based on superpixels and graph cut optimization according to claim 1, wherein in each superpixel adopted in the step (2), one pixel is selected respectively, and these M pixels are utilized Points constitute a new image I _s , proceed as follows:

(2.1) Calculate the gray mean value of each superpixel:

μ μ ((k k)) = = \frac{11}{{m m}_{k k}} {Σ Σ}_{j j = = 11}^{{m m}_{k k}} {sp sp}_{k k} ((j j)),,

Among them, μ(k) is the gray mean value of the kth superpixel sp _k , and m _k represents the number of pixels in the superpixel sp _k ;

(2.2) Select the pixel point corresponding to the smallest value of |sp _k (j)-μ(k)| in the superpixel sp _k , k=1,2,...,M;

(2.3) Use the selected pixels to form a new image I _s .

4. the level set image segmentation method based on superpixels and graph cut optimization according to claim 1, wherein in the step (4), the discrete expression of the level set energy functional is carried out as follows:

(4.1) Define x _p as the binary variable of each pixel in the new image I _s , Among them, p = (x, y) ∈ Ω, Ω represents the image domain;

(4.2) Solve the discrete form of the smooth term of the level set energy functional

{E E.}_{smooth smooth}^{d d} = = μ μ \underset{p p,, q q}{Σ Σ} (({x x}_{p p} ((11 - - {x x}_{q q})) + + {x x}_{q q} ((11 - - {x x}_{p p})))) {w w}_{pq pq},,

Among them, μ is a non-negative parameter, x _p and x _q are binary variables of two different pixel points p and q in the new image I _s respectively, if p and q are connected by axes, then the weight w _pq =1, if p and q are diagonally connected, then the weight

(4.3) Solve the discrete form of the data items of the level set energy functional

{E E.}_{data data}^{d d} = = {λ λ}_{11}^{' '} \underset{p p}{Σ Σ} {| | {I I}_{s the s} - - {c c}_{in in} | |}^{22} {x x}_{p p} + + {λ λ}_{22}^{' '} \underset{p p}{Σ Σ} {| | {I I}_{s the s} - - {c c}_{out out} | |}^{22} ((11 - - {x x}_{p p})),,

Among them, λ′ ₁ and λ′ ₂ are the discrete forms of data items respectively The coefficients of the first item and the second item, I _s represent the new image, c _in and c _out are the mean value of the inner gray level and the mean value of the outer gray level of the closed curve, and the calculation formulas are respectively

{c c}_{in in} = = \frac{{Σ Σ}_{p p} {I I}_{s the s} {x x}_{p p}}{{Σ Σ}_{p p} {x x}_{p p}},, {c c}_{out out} = = \frac{{Σ Σ}_{p p} {I I}_{s the s} ((11 - - {x x}_{p p}))}{{Σ Σ}_{p p} ((11 - - {x x}_{p p}))},,

(4.4) The discrete form of the smooth term of the comprehensive level set energy functional E _CV and the discrete form of the data item Get E _CV discrete form:

\begin{matrix} {E E.}_{CV cv}^{d d} = = {E E.}_{smooth smooth}^{d d} + + {E E.}_{data data}^{d d} \\ = = μ μ \underset{p p,, q q}{Σ Σ} {w w}_{pq pq} (({x x}_{p p} ((11 - - {x x}_{q q})) + + {x x}_{q q} ((11 - - {x x}_{p p})))) \\ + + {λ λ}_{11}^{' '} \underset{p p}{Σ Σ} {| | {I I}_{s the s} - - {c c}_{in in} | |}^{22} {x x}_{p p} \\ + + {λ λ}_{22}^{' '} \underset{p p}{Σ Σ} {| | {I I}_{s the s} - - {c c}_{out out} | |}^{22} ((11 - - {x x}_{p p})) \end{matrix} . .