CN106780376A - The background image dividing method of partitioning algorithm is detected and combined based on conspicuousness - Google Patents

Abstract

The present invention provides a kind of background image dividing method for being detected based on conspicuousness and combining partitioning algorithm, comprises the following steps：S1, original image is processed, obtain single image Saliency maps and multiple images combine Saliency maps；S2, using multiple images combine Saliency maps in significant characteristics as the common conspicuousness of each single image, prospect is made a distinction with background；S3, background image is split.Can disposably by the common foreground segmentation in multiple images out, segmentation adhesion of the normal segmentation method appeared in cutting procedure can be well solved the problems, such as, to reach the purpose of accurate segmentation；And unsupervised conspicuousness detection algorithm is used, this method is realized the automation of cutting procedure again.

Description

The background image dividing method of partitioning algorithm is detected and combined based on conspicuousness

Technical field

The present invention relates to technical field of image processing, detected based on conspicuousness more particularly, to one kind and joint segmentation The background image dividing method of algorithm.

Background technology

Disease geo-radar image identification includes image preprocessing, image segmentation, feature extraction and pattern-recognition.Wherein, image segmentation It is one of committed step, segmentation precision directly affects the reliability of feature extraction and the accuracy of pattern-recognition.Classical segmentation Method has threshold method, rim detection, the dividing method based on statistical-simulation spectrometry, the dividing method based on artificial neural network Deng.

Every kind of dividing method is suitable for each specific situation in the prior art.As Threshold segmentation calculates simple, computing Efficiency is higher, speed is fast, only considers gray value, space characteristics is not considered, to noise-sensitive；After threshold value determines, by threshold value and picture The gray value of vegetarian refreshments is compared one by one, and pixel segmentation can be carried out to each pixel-parallel, and the result of segmentation directly gives figure As region；However, Threshold segmentation is only applicable to the strong image of preceding background contrasts, due in practical situations both, object background and The contrast of prospect is different from everywhere in the middle of image, therefore is difficult have a threshold value for determination.Edge-detected image gray scale Level or the place of structural mutation are detected, so that it is determined that edge distribution, commonly use Roberts operators, Prewitt operators and The first order differential operators such as Sobel operators and the Second Order Differential Operator such as Laplace operators and Kirsh operators carry out rim detection, but It is usually applicable only to the simple image of small noise characteristic.Therefore, classical dividing method also has certain limitation with deficiency.

Under natural conditions, complex background, weather illumination all can cause large effect to the image quality of image, easily make Display foreground and background contrasts are not strong, change very big between different images, cause to use conventional method segmentation effect not good, suitable Answering property is not high.At present, and in the absence of a kind of general dividing method, the image under all features, all situations can be applied to.

It is the method for a kind of new solution segmentation problem for occurring in recent years based on the dividing method that figure cuts, splits in treatment Problem has universality, and segmentation effect is preferable.But in the case where display foreground and background contrasts are not high, segmentation result holds Easily there is adhesion, influence segmentation precision.Some scholars advise being detected from multiple images with common objective or common prospect To strong feature, help to distinguish foreground and background, this kind of method is referred to as joint split plot design.Joint segmentation typically first uses horse to image Er Kefu random fields are modeled, and construct energy equation, solution are then optimized to it with figure segmentation method, to realize dividing segmentation Cut treatment.

The content of the invention

The present invention provide it is a kind of overcome above mentioned problem or solve the above problems at least in part based on conspicuousness detect With the background image dividing method of joint partitioning algorithm, normal segmentation method can be well solved appeared in cutting procedure Split the problem of adhesion, to reach the purpose of accurate segmentation；Using unsupervised conspicuousness detection algorithm, enable this method real again The automation of existing cutting procedure.

According to an aspect of the present invention, there is provided a kind of background image dividing method, comprise the following steps：

S1, original image is processed, obtain single image Saliency maps and multiple images combine Saliency maps；

S2, using multiple images combine Saliency maps in significant characteristics as each single image common conspicuousness, Prospect is made a distinction with background；

S3, background image is split.

Used as preferred, the step S1 is specifically included：By non-supervisory common conspicuousness algorithm to image at Reason, obtains every single image Saliency maps of image and multiple images combine Saliency maps.

Used as preferred, the step S2 includes：

S21, the image optimal scheme for a group with common conspicuousness are marked, and are after marking using Markov random field Image set up one joint segmentation energy equation；

S22, the energy equation by setting up, by the local feature picture with common conspicuousness in single image Saliency maps Element mark is that rest of pixels is labeled as background.

Used as preferred, the step S21 is specifically included：Saliency maps are combined to multiple images by mixed Gauss model Difference with single image Saliency maps is modeled, and constructs as the global restriction of Markov random field model Combine the energy equation of segmentation.

As preferred, in the step s 21, for image I={ I₁,…,I_i,…,I_N, the energy of the joint segmentation Equation is：

E (S)=E_A(S)+E_i(S)

In formula, E_A(S) it is single image internal energy term, E_i(S) it is global energy between image；It is image I_iIn Common significant performance quantifier,It is image I_iIn smoothness energy, η is common significant performance quantifierWeight；P () represents gaussian probability distribution,It is image I_iThe feature of middle pixel k or j,It is image I_iMiddle kth pixel,WithTable Diagram is as I_iThe mixed Gauss model parameter of foreground and background, Θ_comRepresent and the mixed of common objective is constituted by the prospect of all images Close Gauss model parameter.

Used as preferred, the step S22 is specifically included：By energy equation, will there is common conspicuousness in image Local feature pixel is labeled as prospect with internal energy term, in labeling process, by the data item in internal energy term, uses non-prison The conspicuousness detection algorithm superintended and directed extracts every Saliency maps of blade in this group of image, and the smooth item in internal energy term is then roused Encourage the regional area distribution with similar features value on image and obtain uniformity mark, at the same time, weighed by global energy The prospect of each image and the difference of common objective, force the prospect of all images consistent with the common objective of this group of image.

As preferred, in the step S22, when scape is identified with background before dispensing, encourage that there is similar spy on image The regional area distribution uniformity of value indicative is identified, and the smoothed energy is：

In formula, [f] is indicator function, and for prediction f correctly or incorrectly, its value is 1 or is 0,It is image I_i The feature of middle pixel k or j, N is the neighborhood in image, and β is scale coefficient, can be passed throughObtain,<·>Table Diagram is as I_iDesired value, as image I_iWhen middle adjacent pixel is labeled with different labels in Markov random field model, Discontinuously punishment will be made to this.

As preferred, in the step S3, algorithm and iteration are cut by standard drawing so that energy function minimum, realizes Segmentation to entire set of image.

Used as preferred, the step S3 is specifically included：

S31, the color data collection modeling to image so that image is in RGB color by pixelConstitute；

S32, in figure cuts iteration is performed, in the way of iteration energy minimization, background image split.

Compared with prior art, the beneficial effects of the present invention are：The present invention is examined with non-supervisory common conspicuousness first Method of determining and calculating is every image common Saliency maps of generation in one group of image, and it is random that these Saliency maps are used to construction Markov Internal image energy function in.Using mixed Gauss model to the common objective of this group of image whole Saliency maps and individual The difference of image Leaf is modeled, and a new global restriction as Markov random field optimal model is gone Construction global energy；Cut algorithm and iteration finally by standard drawing so that energy function minimize, to realize to cotton seedling leaf The segmentation of picture.Can disposably by the common foreground segmentation in multiple images out, can well solve normal segmentation The problem of segmentation adhesion of the method appeared in cutting procedure, to reach the purpose of accurate segmentation；And use unsupervised notable Property detection algorithm, enables this method realize the automation of cutting procedure again.

Brief description of the drawings

Fig. 1 is the method flow block diagram of the embodiment of the present invention；

Fig. 2 be the embodiment of the present invention in method be used for cotton background image split schematic flow sheet；

Fig. 3 is every common Saliency maps of image generation in cotton original image in the embodiment of the present invention；

Fig. 4 is cotton original image segmentation result schematic diagram in the embodiment of the present invention.

Specific embodiment

With reference to the accompanying drawings and examples, specific embodiment of the invention is described in further detail.Hereinafter implement Example is not limited to the scope of the present invention for illustrating the present invention.

Fig. 1 shows a kind of flow chart of the background image dividing method for being detected based on conspicuousness and combining partitioning algorithm, Comprise the following steps：

S1, original image is processed, obtain single image Saliency maps and multiple images combine Saliency maps；

S2, using multiple images combine Saliency maps in significant characteristics as each single image common conspicuousness, Prospect is made a distinction with background；

S3, background image is split.

One group of original image is given for I={ I₁,…,I_i,…,I_N, pixel byIt is represented, N_iRepresent I_iImage I lattice,Represent I_iPixel in imageNormalization position.Image wherein I_iThe binary of image is designated It is I_iIn k-th mark of pixel,For 1 represents prospect, it is 0 and is expressed as background, n_iIt is I_i's Total pixel number.Wherein image inside cluster numbers are K₁, cluster numbers are K between image₂。

The step S1 is processed image by non-supervisory common conspicuousness algorithm, obtains every image individual Image saliency map and multiple images combine Saliency maps.

In this embodiment, the step S1 is specifically included：

S11, by image inside cluster segmentation in original image into K₁Class, calculates each by common conspicuousness detection method The contrast clue and spatial cues of class, obtain single image Saliency maps；

S12, by cluster segmentation between image into K₂Class, the contrast line of each class is calculated by common conspicuousness detection method Rope, spatial cues and correlation clue, with reference to single image Saliency maps, obtain multiple images joint Saliency maps.

Specifically, the step S11 is specifically included：

S111, first to this group of original image I={ I₁,…,I_i,…,I_NIn N width image clusterings be divided into K₁Class, obtains To K₁Phylogenetic groupPhylogenetic groupBy one group of D dimensional vectorTo represent.μ^kRepresent and cluster C^kAssociated is poly- Class center, function ψ：L²→ { 1 ..., K } and pixelIndexed with focusing onIt is associated；

S112, the contrast clue and spatial cues that calculate each class.

Wherein contrast clue is calculated as follows：

Wherein, L₂Norm is used to calculate the distance of feature space, nⁱRepresent cluster CⁱPixel count, N represents all images Pixel count.

Wherein spatial cues are calculated as follows：

Wherein, δ () is Kronecker delta functions (the kronecker δ function), o_iRepresent I_iThe center of image, it is high This core N () is used to calculate pixelWith picture centre o_iEuclidean distance, variances sigma²It is the normalization radius of image.Normalization Coefficient n^kIt is cluster C^kPixel count.Same, the spatial cues ω with single image model^sThe position between cluster rank is represented, is represented Global center deviation between multiple image；

Circulation is performed to more than based on the contrast clue calculating process on the basis of cluster and spatial cues calculating process to refer to Order, until by original image I={ I₁,…,I_i,…,I_NIn all images cluster be all disposed by stop computing.

S113, contrast clue calculating process and spatial cues calculating process are combined by being defined as below, so that To cluster K₁Joint significance collection：

Wherein, w_iRepresent conspicuousness clue.Before conspicuousness clue is combined, our normalizations each clue figures makes Balanced Gaussian function is removed with the score distribution of all clusters；

S114, calculate joint conspicuousness level of constellation value, there is provided discrete distribution, then smooth each pixel connection The value of conspicuousness is closed, the significant likelihood of pixel x is satisfied with a C of Gaussian Profile N^kAggregated pattern：

S115, single Saliency maps are obtained with equation below, so as to obtain the joint conspicuousness of set of pixels level：

S116, the process execution recursion instruction based on the acquisition joint Saliency maps in pixel basis, Zhi Daozhi to more than To by original image I={ I₁,…,I_i,…,I_NIn all images pixel be all disposed by stop computing.

In the present embodiment, the step S12 is specifically included：

S121, to this group of original image I={ I₁,…,I_i,…,I_NIn N width image clusterings be divided into K₂Class, obtains K₂ Phylogenetic groupPhylogenetic groupBy one group of D dimensional vectorTo represent.μ^kRepresent and cluster C^kIn associated cluster The heart, function ψ：L²→ { 1 ..., K } and pixelIndexed with focusing onIt is associated；

S122, the contrast clue ω for calculating each class^c(k), spatial cues ω^s(k) and correlation clue ω^d(k)；

Wherein contrast clue is calculated as follows：

Wherein, L₂Norm is used to calculate the distance of feature space, nⁱRepresent cluster CⁱPixel count, N represents all images Pixel count.

Wherein spatial cues are calculated as follows：

Wherein, δ () is Kronecker delta functions, o_iRepresent I_iThe center of image, Gaussian kernel N () is based on Calculate pixelWith picture centre o_iEuclidean distance, variances sigma²It is the normalization radius of image.Normalization coefficient n^kIt is cluster C^k's Pixel count.Same, the spatial cues ω with single image model^sThe position between cluster rank is represented, is represented between multiple image Global center deviation.

Wherein correlation clue is calculated as follows：

First, M-bin histogramFor describing to cluster C in N images^kDistribution：

n^kIt is cluster C^kPixel count, force

Then correlation clue is calculated as follows：

Wherein, var (q^k) have recorded cluster C^kHistogram q^kVariation, high correspondence clue collection group representation, the picture of the cluster Element is evenly distributed on every image.

S123, execution circulation contrast clue calculating process, spatial cues calculating process and correlation clue were calculated Journey, until by original image I={ I₁,…,I_i,…,I_NIn the pixel cluster of all images be all disposed；

S124, with reference to single image Saliency maps and three kinds of clue formulasCarry out at computing Reason；

S125, formulaThe final joint conspicuousness of smooth gained Figure.

The process based on the final joint Saliency maps of the smooth gained in pixel basis performs recursion instruction to more than, directly To until by original image I={ I₁,…,I_i,…,I_NIn all images pixel be all disposed by stop computing.

In the present embodiment, the step S2 includes：

S21, the image optimal scheme for a group with common conspicuousness are marked, and are after marking using Markov random field Image set up one joint segmentation energy equation；

S22, the energy equation by setting up, by the local feature picture with common conspicuousness in single image Saliency maps Element mark is that rest of pixels is labeled as background.

S21 is specifically included：Saliency maps and single image Saliency maps are combined to multiple images by mixed Gauss model Difference be modeled, and as Markov random field model global restriction come tectonic syntaxis segmentation energy side Journey.

It is the image optimal scheme mark with common conspicuousness, by Markov random field for these images set up one The individual joint for being added and being obtained by global energy between single image internal energy term and image splits energy equation：

E (S)=E_A(S)+E_i(S)

In formula, E_A(S) it is single image internal energy term, E_i(S) it is global energy between image；It is image I_iIn Common significant performance quantifier,It is image I_iIn smoothness energy, η is common significant performance quantifierWeight；P () represents gaussian probability distribution,It is image I_iThe feature of middle pixel k or j,It is image I_iMiddle kth pixel,With Represent image I_iGMM (Gaussian Mixture Model, gauss hybrid models) model parameter, Θ of foreground and background_comTable Show the GMM model parameter that common objective is made up of the prospect of all images.

The step S22 is specifically included：Pass through set up energy equation, will there is common conspicuousness in Saliency maps picture Local feature pixel be labeled as prospect, rest of pixels be labeled as background；

By energy equation, by the local feature pixel with common conspicuousness in image with internal energy term labeled as preceding Scape, in labeling process, by the data item in common conspicuousness model internal energy term, is detected using non-supervisory conspicuousness Algorithm extracts every Saliency maps of blade in this group of image, and the smooth item in internal energy term then encourages have phase on image Uniformity mark is obtained like the regional area distribution of characteristic value, so that the foreground features of prominent Saliency maps, make foreground features more Substantially；At the same time, the global energy for being defined by gauss hybrid models, weighs the prospect of each image and the difference of common objective It is different, force the prospect of all images consistent with the common objective of this group of image.

In allocation identification, the regional area distribution uniformity mark on image with similar features value is encouraged, it is described flat Sliding energy term is：

In formula, [f] is indicator function, and for prediction f correctly or incorrectly, its value is 1 or is 0,It is image I_i The feature of middle pixel k or j, N is the neighborhood in image, and β is scale coefficient, can be passed throughObtain,<·>Represent Image I_iDesired value, as image I_iWhen middle adjacent pixel is labeled with different labels in MRF models,Will be to this discontinuous Make punishment.

In the present embodiment, the step S3 is specifically included：Algorithm is cut to being optimized to energy equation using standard drawing, The difference of each image prospect and common objective is weighed by the global energy constructed by image feature data, using iteration side Method, for each segmentation result, every image background is polymerized to using K-means clustering methodsThe quantity K of middle Gaussian kernel_b, The expectation and variance of the central value and variance of each class as corresponding Gaussian kernel are calculated, while in an iterative process by all figures The prospect of picture tries to achieve Θ as common objective using same method_comMiddle K_cGroup GMM parameters, finally, are scratched at figure using edge Reason image.

Fig. 2 is applied to the schematic flow sheet in the segmentation of cotton background image for the method for the present invention, is total to by non-supervisory With conspicuousness detection algorithm, using CIE Lab colors (CIE Lab：L * component in color space represents the brightness of pixel；A tables Show the scope from red to green；B is represented from yellow to blue scope) and it is next with the Gabor filter that 8 orientation are responded Represent characteristic vector, one scale of 1 bandwidth of selection and extraction.By combining 8 textural characteristics in orientation, (8 orientation are responded Gabor wavelet energy：Deflection Θ according to wave filter is different, is divided into eight orientation：45 degree, 90 degree, 135 degree, 180 degree, 225 degree, 270 degree, 315 degree, 360 degree), calculate the image size of Gabor filter.K averages are applied into two-layer cluster to work as In.Single image is directed to, the cluster numbers of common conspicuousness detection algorithm are set to 6, and for multiple images, cluster numbers are then set to Min { 3M, 20 }, wherein M represents picture number, obtains one group of gray-scale map with common conspicuousness, as shown in figure 3, be respectively locating Before reason after (a, b, c, d) and treatment common conspicuousness gray-scale map (e, f, g, h) schematic diagram.

Joint segmentation to the common objective optimal scheme mark in N figures, by every internal energy term of image and External energy term sum between image carries out optimizing and obtains.The binary of note N width image distribution is designated S={ s₁,…,s_N, lead to Cross to energy equation E (S)=E_A(S)+E_i(S) minimum optimizing is carried out, optimal solution S can be obtained^*, so as to obtain optimum segmentation result.

Pass through set up energy equation, the local feature pixel with common conspicuousness in Saliency maps picture is labeled as Prospect, rest of pixels is labeled as background；

Using alternative manner, minimum treatment is carried out to energy equation, to realize the segmentation to entire set of image.

First, to the color data collection z of image_nModeling so that image is in RGB color by pixelConstitute.

Iteration is performed in figure cuts, in this mode of iteration energy minimization, instead of previous disposable partitioning algorithm.

First, it is directed to the T of initial pictures_UEach pixel in regionIt is pixel one gauss hybrid models of establishment Inscape：

Then, learning gauss hybrid models coefficient is worked as from data set z：

Segmentation equation is then set up, is understood using minimal cut place to go and determined：

And then brought into operation from newly from the first step, until set restrains；

Using edge FIG pull handle image, result is finally given as shown in figure 4, for one group of original image box segmentation result shows It is intended to (figure is the gray level image of original graph and segmentation result).

In sum, the present invention is every image life in one group of image first with non-supervisory common conspicuousness detection algorithm Into common Saliency maps, these Saliency maps are used to construct the internal image energy function in markov random file.Using Mixed Gauss model is modeled to the common objective of this group of image whole Saliency maps with the difference of single image Leaf, will It removes construction global energy as a new global restriction of Markov random field optimal model；Finally by standard Figure cut algorithm and iteration so that energy function minimize, to realize the segmentation to cotton seedling leaf image.Can disposably will be many Common foreground segmentation in image out, can well solve segmentation of the normal segmentation method appeared in cutting procedure and glue Problem even, to reach the purpose of accurate segmentation；And unsupervised conspicuousness detection algorithm is used, this method is realized point again Cut the automation of process.

Finally, the present processes are only preferably embodiment, are not intended to limit the scope of the present invention.It is all Within the spirit and principles in the present invention, any modification, equivalent substitution and improvements made etc. should be included in protection of the invention Within the scope of.

Claims (9)

1. a kind of background image dividing method, it is characterised in that comprise the following steps：

S1, original image is processed, obtain single image Saliency maps and multiple images combine Saliency maps；

S2, using multiple images combine Saliency maps in significant characteristics as each single image common conspicuousness, to preceding Scape makes a distinction with background；

S3, background image is split.

2. background image dividing method according to claim 1, it is characterised in that the step S1 is specifically included：Pass through Non-supervisory common conspicuousness algorithm is processed image, obtains every single image Saliency maps of image and multiple images Joint Saliency maps.

3. background image dividing method according to claim 1, it is characterised in that the step S2 includes：

S21, the image optimal scheme for a group with common conspicuousness are marked, and are the figure after mark using Markov random field As setting up an energy equation for joint segmentation；

S22, the energy equation by setting up, by the local feature pixel mark with common conspicuousness in single image Saliency maps Prospect is designated as, rest of pixels is labeled as background.

4. background image dividing method according to claim 3, it is characterised in that the step S21 is specifically included：Pass through Mixed Gauss model is modeled to multiple images joint Saliency maps with the difference of single image Saliency maps, and as The global restriction of Markov random field model carrys out the energy equation of tectonic syntaxis segmentation.

5. background image dividing method according to claim 4, it is characterised in that in the step s 21, for image I= {I₁,…,I_i,…,I_N, the energy equation of the joint segmentation is：

E (S)=E_A(S)+E_i(S)

$E_i\left(S\right)=\underset{i=1}{\overset{N}{\&Sigma;}}\&lsqb;{\mathrm{\&eta;E}}_i^C\left(S\right)+E_i^S\left(S\right)\&rsqb;$

$E_A\left(S\right)={\&Sigma;}_i{\&Sigma;}_k-log\&lsqb;s_k^{i}P\left(v_k^i\right|{\mathrm{\&Theta;}}_{com})+(1-s_k^i)P\left(v_k^i\right|{\mathrm{\&Theta;}}_b^i)\&rsqb;$

In formula, E_A(S) it is single image internal energy term, E_i(S) it is global energy between image；E_i ^CIt is image I_iIn it is common Significant performance quantifier,It is image I_iIn smoothness energy, η is common significant performance quantifier E_i ^CWeight；P () is represented Gaussian probability is distributed,It is image I_iThe feature of middle pixel k or j,It is image I_iMiddle kth pixel,WithRepresent image I_i The mixed Gauss model parameter of foreground and background, Θ_comExpression is made up of the mixed Gaussian mould of common objective the prospect of all images Shape parameter.

6. background image dividing method according to claim 5, it is characterised in that the step S22 is specifically included：Pass through Energy equation, prospect, labeling process are labeled as by the local feature pixel with common conspicuousness in image with internal energy term In, by the data item in internal energy term, every blade in this group of image is extracted using non-supervisory conspicuousness detection algorithm Saliency maps, the smooth item in internal energy term then encourage on image with similar features value regional area distribution obtain one Cause property mark, at the same time, the prospect and the difference of common objective of each image is weighed by global energy, forces all images Prospect it is consistent with the common objective of this group of image.

7. background image dividing method according to claim 6, it is characterised in that in the step S22, before dispensing scape When being identified with background, the regional area distribution uniformity mark on image with similar features value, the smoothed energy are encouraged For：

$E_i^S\left(S_i\right)=\underset{(k,j)\&Element;N}{\&Sigma;}\&lsqb;S_k^i\&NotEqual;S_j^i\&rsqb;\exp (-\mathrm{\&beta;}||v_k^i-v_j^i|\left|\right)$

In formula, [f] is indicator function, and for prediction f correctly or incorrectly, its value is 1 or is 0,It is image I_iMiddle picture The feature of plain k or j, N is the neighborhood in image, and β is scale coefficient, can be passed throughObtain,<·>Represent image I_iDesired value, as image I_iWhen middle adjacent pixel is labeled with different labels in Markov random field model,Will be to this Plant and discontinuously make punishment.

8. background image dividing method according to claim 1, it is characterised in that in the step S3, by standard drawing Algorithm and iteration are cut so that energy function minimum, realizes the segmentation to entire set of image.

9. background image dividing method according to claim 8, it is characterised in that the step S3 is specifically included：

S31, the color data collection modeling to image so that image is in RGB color by pixelConstitute；

S32, in figure cuts iteration is performed, in the way of iteration energy minimization, background image split.