CN107818566A - A kind of image partition method based on partial structurtes information around random walk combination pixel - Google Patents

A kind of image partition method based on partial structurtes information around random walk combination pixel Download PDF

Info

Publication number
CN107818566A
CN107818566A CN201711019238.9A CN201711019238A CN107818566A CN 107818566 A CN107818566 A CN 107818566A CN 201711019238 A CN201711019238 A CN 201711019238A CN 107818566 A CN107818566 A CN 107818566A
Authority
CN
China
Prior art keywords
image
pixel
node
random walk
segmentation
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201711019238.9A
Other languages
Chinese (zh)
Inventor
蒋庆
郑杰文
宋振华
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Sun Yat Sen University
National Sun Yat Sen University
Original Assignee
National Sun Yat Sen University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by National Sun Yat Sen University filed Critical National Sun Yat Sen University
Priority to CN201711019238.9A priority Critical patent/CN107818566A/en
Publication of CN107818566A publication Critical patent/CN107818566A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/143—Segmentation; Edge detection involving probabilistic approaches, e.g. Markov random field [MRF] modelling
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/194—Segmentation; Edge detection involving foreground-background segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10004—Still image; Photographic image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10024—Color image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20024—Filtering details
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20076—Probabilistic image processing

Abstract

A kind of image partition method based on partial structurtes information around random walk combination pixel of the present invention, the present invention is good to image segmentation, and accuracy is high, especially has higher accuracy to the segmentation problem containing partial structurtes complicated image.The invention belongs to image procossing, image segmentation algorithm field.General pattern dividing method is solved because local message is complicated and splits misalignment problem.The present invention is applicable not only to the segmentation of the general clear-cut obvious image of structure, the segmentation to local structural information complicated image is also applied for, to also having good result containing elongated portion cutting object.

Description

A kind of image segmentation based on partial structurtes information around random walk combination pixel Method
Technical field
Present invention relates particularly to a kind of image segmentation side based on partial structurtes information around random walk combination pixel Method.
Background technology
The shortcomings of the prior art is：
(1) existing image segmentation algorithm, segmentation limitation is big, and the scope of application is not wide, to the segmentation effect of different objects It is not good enough.Such as existing random walk image segmentation algorithm is only suitable for the image that cutting object is obvious, clear-cut, and for tool There is image segmentation in disorder, complicated, that marginal texture texture information is complicated bad.
(2) existing random walk image segmentation algorithm, the image enriched to local structural information, object bounds feature are not When the complicated image of obvious image and texture information is split, the segmentation effect of its image segmentation is bad, and segmentation result is past Toward inaccuracy.
(3) existing image segmentation algorithm, treat in segmentation figure picture containing elongated and the object to be split of end points The segmentation result of image is inaccurate, can not meet the image segmentation for including elongated portion.Utilize existing image partition method Treat when being split in segmentation figure picture containing the cutting object of elongated part, the phenomenon of less divided often occur.
Existing algorithm causes the reason for above deficiency to have：
(1) traditional Random Walk Algorithm only calculates no label pixel point to the probability of label node.It is existing random Migration image segmentation algorithm does not consider addition assisted tag, so that when calculating each pixel arrival label point probability, pixel The foundation of the tag along sort of point is not abundant enough, and the effect for causing image to be split is bad, and robustness is not high.Especially contain in segmentation During the image of complicated elongated cutting object, traditional Random Walk Algorithm easily produces the phenomenon of less divided.
(2) existing image Segmentation Technology all only emphasize and consider in Graph Theory the pixel of image color and Half-tone information, and the partial structurtes information among image around pixel is not considered, which results in contain abundant structure in segmentation During the texture image of information, segmentation effect is bad.There is the cutting object of complicated partial structurtes, its point particularly with contour edge It is often inaccurate to cut result.The Study Of Segmentation Of Textured Images result obtained using existing random walk image Segmentation Technology is not good enough.
(3) when cutting object contains elongated portion, existing random walk image Segmentation Technology partitioning algorithm seed mark Label set the requirement for failing to meet its Accurate Segmentation.And existing random walk image Segmentation Technology is utilizing Di Li Crays border Problem calculate the pixel node without label to label seed point transition probability when, only only account for the class label of single form Node, fail to be apparent to describe pixel classification foundation exactly.Pixel, which is calculated, which results in image finally distributes label Probability is not accurate enough, contains the image of elongated portion particularly with cutting object, traditional random walk image segmentation algorithm because Elongated portion surrounding tags node for cutting object is limited and can produce the result of less divided.
In summary, existing random walk image segmentation algorithm fails to meet, edge contour structure complicated to structural information The requirement of image Accurate Segmentation comprising elongated portion in complicated and cutting object.
Its reason is summarised as：
1st, the tag along sort Seed label form of the former setting of traditional random walk image segmentation algorithm is single, pixel The foundation of tag along sort is not abundant enough, and the result for causing image to be split is not accurate enough, and specific manifestation can not meet to cutting object Accurate segmentation requirement containing elongated portion；
2nd, existing image Segmentation Technology is when graph theory describes the similarity degree pixel node, only only in accordance with single Pixel color or half-tone information, the structural information on its pixel periphery, therefore existing random walk image are not taken into full account Dividing method can not the accurate baroque image of segmenting edge.
The content of the invention
It is an object of the invention to provide a kind of high accuracy, the image point for being good at handling local labyrinth image Algorithm is cut, to solve the technical problem in background technology.
To realize object defined above, the present invention provides following technical scheme：
A kind of image partition method based on partial structurtes information around random walk combination pixel, specific algorithm flow Comprise the following steps：
S1：Original image is inputted, program reads image to be split；
S2：Auxiliary scribble is classified with adding tag along sort and adding assisted tag node and link in image to be split Label node, the tag along sort are divided into background label and prospect label；
S3：Using pixel information and structure tensor, form novel local structural information and define in graph theory between pixel Similarity weight function wij, to represent the similarity degree between them；
S4：With Markov property addition assisted tag node image in set random walk person pixel node it Between transition probability；
S5：The transition probability between random walk person's pixel is solved using the Di Li Crays boundary function of combination；
S6：Random walk person is calculated from the node without label into the probability in all label nodes using formula Maximum；To reach label node maximum probability as distribution tag along sort that principle is each pixel；
S7：It is recycled to untill whether there is label node all distributes to obtain tag along sort；Each pixel distribution is divided Class label completes final segmentation to image.
The image that will split is mapped to a non-directed graph G=(V, E), by node v ∈ V and borderGroup Into.Each node viRepresentative image pixel xi, eijRepresent two vertex vsiAnd vjLink, wij∈ W are represented between two summits Similarity degree, here wij>=0, and wij≥wji。
The degree of similarity represents that weighting function equation is as follows with dimensional Gaussian weight function：
Wherein IiAnd IjIt is the rgb value or gray value of the pixel of image, and JiAnd JjRepresent the structure of partial structurtes information The scaling vector of tensor.
JiAnd JjRepresent the scaling vector of the structure tensor of partial structurtes information.In formula, | | Ji-Jj| | represent standard L2Europe Formula norm.
Wherein for two-dimensional image I: Ω → R2, arbitrarily certain gradient vector for putting pixel is in the picture Then the tensor of the point is：
And from detG=0, G only has a nonzero eigenvalue, and therefore, tensor can only describe the one-dimentional structure of pixel And direction, and the multidimensional information around pixel can not be described；
In order to embody partial structurtes information, filtering technique can be used to carry out matrix field data smooth, by filtered popin Tensor after cunning is defined as structure tensor, and its expression formula is：
Wherein hσFor the adaptive-filtering function of definition, its form is：
In formula, r is spatial frequency, and σ is variances sigma/(λ of gaussian kernel function1+λ2) size determine the work of filter function With scope, if σ/(λ1+λ2) larger, illustrate that characteristic value sum is larger, partial structurtes are uniform domain, then using larger filtering Scope；If conversely, σ/(λ1+λ2) smaller, illustrate that partial structurtes are edge or angle point region, then using less filter range, this Outside, parameter θ represents the edge direction at (x, y) place, in the same direction with the characteristic vector corresponding to smaller characteristic value；
So, the structure tensor T based on new adaptive anisotropic filtering function can be written as：
In pretreatment stage, first using anisotropic filtering function mentioned above, image is transformed into structure tensor Space, the mark and uniformity of the counter structure tensor of each pixel are then asked for, obtain scaling vector J (x, y)；
Utilize the mark trG of structure tensorσ=λ1+λ2To describe the intensity of localized variation；Define cohGσ=(λ1-λ2)2To carve Draw the uniformity of partial structurtes；
Define J (x, y)=[trGσcohGσ], scaling vector J (x, y) illustrates the adjacent structural information of pixel, has more Good picture structure ability to express, the expression-form that can obtain connection weight between the pixel needed for Random Walk Algorithm are：
Wherein IiAnd IjIt is the rgb value or gray value of the pixel of image, and JiAnd JjRepresent the structure of partial structurtes information The scaling vector of tensor.
Constant is arranged to ρ=10-7, control parameter is arranged toControl parameter σ1, σ2Pass through experiment Different values is taken to be contrasted, actual result can be set, when in image split-run testWhen, segmentation effect It is most stable.
It is the non-directed graph containing assisted tag node is added that patent of the present invention, which expands the non-directed graph of image, as shown in Fig. 2 nothing The node V of labelUWith the seed node V for having labelM, wherein VU∪VM=V, S in patent of the present inventionMIt is assisted tag point with Δ, Then compare in patent of the present invention and solve each node to the similarity probabilities of label node, can try to achieve to each picture in image The labeling of vegetarian refreshments, it is the result of proprietary algorithms image segmentation of the present invention.
The sub- Markov Transition Probabilities property definition：
Q is expressed as transition probability of no label node on V；By adding assisted tag point Δ and SM, according to Ma Erke The characteristic of husband's transition probability, there is following property：The V ∪ (Δ) on the non-directed graph of addition assisted tag, set and define q (Δ, Δ) =1 HeThen q (i, Δ) represents the probability that a random walk person leaves from figure G.
As shown in Fig. 2 adding two kinds of assisted tag nodes in patent of the present invention, one kind is linked without label seed node , node Δ is referred to as eliminated, when random walk person reaches these nodes, it will link on the nodes, corresponding general Rate can eliminate, that is to say, that a random walk person will not reach such node；Another kind is to link label node, is referred to as protected Stay nodeRandom walk person will be remained when reaching these nodes, these nodes are referred to as destination node.We are also fixed Adopted SMFor relative to VMReservation node, ciFor the probability left from V interior joints.
The random walk person is from node vi∈ V are in V ∪ { Δ } ∪ SMTransition probability may be defined as：
As random walk person from node vi∈ V, which set out, reaches auxiliary reservation nodeProbability, then have：
Wherein, to gauge method(N=| V | be the quantity of node) vector mark can be expressed as
Wherein,Be withThe N-dimensional row vector of composition；DcIt is with diagonal element ciPair of composition Angular moment battle array, i.e. Dc=diag (c1, c2..., cN), and I is N × N unit matrix；
Transition probability matrix P=[pij]N×NIt is the row normalized matrix of weight matrix W between pixel：
pij=wij/di
Utilize average arrival probabilityThat is a random walk person is from node viSet out and reach the reservation node of label, table Show and distribute to node label lkPossibility probability；Then formula is changed into；
Wherein E=I- (I-Dc)P；
The probability of stabilityCalculation formula it is as follows：
WhereinIt is to work asWhenWith in the case of otherColumn vector, ZkIt is that normalization is normal Number；
Last each node vi=V distribution label result (i.e. the segmentation result of image) is determined by following formula：
Beneficial effect of the present invention：Patent of the present invention is good to image segmentation, and accuracy is high, especially to containing local knot The segmentation problem of structure complicated image has higher accuracy.Patent of the present invention belongs to image procossing, image segmentation algorithm field. General pattern dividing method is solved because local message is complicated and splits misalignment problem.Patent of the present invention is applicable not only to typically The segmentation of the clear-cut obvious image of structure, is also applied for the segmentation to local structural information complicated image, to containing elongate portion Divide cutting object that also there is good result.
(1) patent of the present invention can accurately be split to complicated, mixed and disorderly and complicated contour edge cutting object, Precise results are obtained, the effect than existing random walk image partition method is good.Patent of the present invention can not only meet existing random Segmentation requirement of the migration image segmentation to image clear in structure, can also meet will to the complicated image Accurate Segmentation of contour edge Ask.Patent i.e. of the present invention will solve existing random walk can not be to complicated, mixed and disorderly and complicated contour edge segmentation The problem of object is accurately split.
(2) present invention has the cutting object of abundant partial structurtes information and segmentation effect is good, has high robust and height Stability.Patent of the present invention can be carried out to the obvious image of the clear profile of general structure with the image containing complex texture information Accurate Segmentation, and this method robustness is high, stability is high.Patent i.e. of the present invention, which will solve existing random walk, can split typically The obvious image of profile clear in structure but can not be containing complex texture information image the problem of.Existing random walk to containing The image segmentation of complex texture information is bad, and robustness is not high.
(3) patent of the present invention can solve existing Random Walk Algorithm pair by adding auxiliary note node in graph theory Cutting object contains the inaccurate problem of segmentation of elongated portion.Patent of the present invention can be entered to the cutting object containing elongated portion The accurate segmentation of row.Because patent of the present invention with the addition of assisted tag, solves existing random walk image segmentation algorithm because treating Segmentation figure can not calculate the label allocation probability of elongated portion pixel exactly as elongated portion surrounding tags node is short of, The problem of so as to cause image elongated portion less divided.Therefore patent of the present invention contains the segmentation knot of elongated portion to cutting object Fruit is accurate.
The image accuracy of separation is high, and segmentation effect is good, and image segmentation robustness and stability are high.Local structural information is answered The segmentation that miscellaneous, texture information enriches and cutting object contains elongated portion image has good segmentation effect.
1, patent of the present invention splits skill in terms of splitting with abundant texture information image, than existing random walk image Art is more superior.The image segmentation of patent of the present invention is good, accuracy is high, has the characteristics of high robust and high stability.This Patent of invention is to general pattern and the image with abundant texture information is all with good segmentation result.
2, it is not only obvious to general cutting object compared to other existing random walk image partition methods, patent of the present invention Image accurately split, and the image to be split containing elongated complicated part also can accurately be split, Er Qiefen It is higher to cut result accuracy.
In image segmentation, accuracy is to weigh the standard of image segmentation accuracy, emphasizes segmentation figure picture and graphics standard point Segmentation difference between cutting.Accuracy is higher, represents the image after segmentation closer to Standard Segmentation.Accuracy in the present invention is determined Justice is：
Wherein StpIt is the percentage in the region correctly split；StnIt is the background percentage correctly split；SfpIt is erroneous segmentation Target area percentage；SfnIt is the background percentage of erroneous segmentation.Current Fig. 2 texture images are illustrated：
It can be calculated by above-mentioned formula：
The accuracy of the present invention is 98.83%, and the accuracy of traditional random walk method is 80.62%, specifically be see the table below：
Comparative approach Accuracy (%)
Traditional random walk method 80.62
The present invention 98.83
Brief description of the drawings
The schematic flow sheet of Fig. 1 present invention
Fig. 2 adds auxiliary node schematic diagram.
It is added to the addition prospect label and background label information of interaction in figure in Fig. 3 a on figure for expression.
The result profile with abundant texture information image segmentation is expressed as in Fig. 3 b, fine rule sketches the contours of the wheel of cutting object It is wide.
The segmentation result of the image with abundant texture information is expressed as in Fig. 3 c.
The Standard Segmentation result figure of the manual identification of the image with abundant texture information is expressed as in Fig. 3 d.
It is added to the addition prospect label and background label information of interaction in figure in Fig. 4 a on figure for expression.
The image segmentation result profile with cutting object elongated portion is expressed as in Fig. 4 b, fine rule sketches the contours of cutting object Profile.
The segmentation result with the image of cutting object elongated portion is expressed as in Fig. 4 c.
The Standard Segmentation result figure with the manual identification of the image of cutting object elongated portion is expressed as in Fig. 4 d.
Embodiment
Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clear, complete Site preparation describes.
Refer to Fig. 1-4, the present invention is a kind of partial structurtes information of combination image, a kind of novel interactive of proposition with Machine migration image segmentation algorithm.The patent of invention not only allows for the color and half-tone information of original image slices vegetarian refreshments, it is also considered that The pixel partial structurtes information of image, can more accurately describe the relation between pixel.So that the image point of the present invention It is more extensive to cut algorithm applicability, and robustness is higher.
Patent of the present invention is a kind of new multi-tag, with the addition of assisted tag and combine partial structurtes texture information Interactive random walk image segmentation algorithm.
In patent of the present invention, band is weighed again between image is regarded as a little non-directed graph solves, that is, divides the image into and be converted into The segmentation problem of non-directed graph.The image that will split is mapped to a non-directed graph G=(V, E), by node v ∈ V and borderComposition.Each node viRepresentative image pixel xi, eijRepresent two vertex vsiAnd vjLink, wij∈ W are represented Similarity degree between two summits.Here wij>=0, and wij≥wji。
Pay special attention to, patent of the present invention is characterized in the degree of similarity dimensional Gaussian weight function table between non-directed graph summit Show, weighting function equation is as follows：
Wherein IiAnd IjIt is the rgb value or gray value of the pixel of image, and JiAnd JjRepresent the structure of partial structurtes information The scaling vector of tensor.
According to claims, it is characterised in that JiAnd JjRepresent the scaling vector of the structure tensor of partial structurtes information. In formula, | | Ji-Jj| | represent standard L2European norm.
Wherein for two-dimensional image I：Ω→R2, arbitrarily certain gradient vector for putting pixel is in the picture Then the tensor of the point is
And from det G=0, G only has a nonzero eigenvalue, and therefore, tensor can only describe the one-dimensional knot of pixel Structure and direction, and the multidimensional information around pixel can not be described.In order to embody partial structurtes information, filtering technique pair can be used Matrix field data carries out smooth.Tensor after will be smooth after filtering is defined as structure tensor, and its expression formula is
Wherein hσFor the adaptive-filtering function of definition, its form is
In formula, r is spatial frequency, and σ is variances sigma/(λ of gaussian kernel function1+λ2) size determine the work of filter function With scope, if σ/(λ1+λ2) larger, illustrate that characteristic value sum is larger, partial structurtes are uniform domain, then using larger filtering Scope；If conversely, σ/(λ1+λ2) smaller, it is edge or angle point region to illustrate partial structurtes, then using less filter range.This Outside, parameter θ represents the edge direction at (x, y) place, in the same direction with the characteristic vector corresponding to smaller characteristic value.
So, the structure tensor T based on new adaptive anisotropic filtering function can be written as：
In pretreatment stage, first using anisotropic filtering function mentioned above, image is transformed into structure tensor Space, the mark and uniformity of the counter structure tensor of each pixel are then asked for, obtain scaling vector J (x, y).It is of the invention special Profit utilizes the mark trG of structure tensorσ=λ1+λ2To describe the intensity of localized variation；Define coh Gσ=(λ1-λ2)2To portray office The uniformity of portion's structure.J (x, y)=[trG defined in patent of the present inventionσcoh Gσ], the scaling vector of patent so of the present invention J (x, y) illustrates the adjacent structural information of pixel, has better image structure representation ability.
The expression-form that connection weight between the pixel needed for Random Walk Algorithm can be obtained is：
Wherein IiAnd IjIt is the rgb value or gray value of the pixel of image, and JiAnd JjRepresent the structure of partial structurtes information The scaling vector of tensor.Constant is arranged to ρ=10-7, control parameter is arranged toControl parameter σ1, σ2It is logical Crossing experiment takes different values to be contrasted, and actual result can be set, when in image split-run testWhen, point It is most stable to cut effect.
Next, patent of the present invention expands the non-directed graph of image the non-directed graph for the assisted tag node containing addition, such as Fig. 1 It is shown, the node V of no labelUWith the seed node V for having labelM, wherein VU∪VM=V, S in patent of the present inventionMIt is auxiliary with Δ Label point, then compare in patent of the present invention and solve each node to the similarity probabilities of label node, can try to achieve to image In each pixel labeling, be proprietary algorithms image segmentation of the present invention result.
Patent of the present invention, it is characterized in that representing that q is the transition probability without label node on V.Turned according to sub- Markov Move the property definition of probability：
Patent of the present invention, it is characterised in that by adding assisted tag point Δ and SM, according to Markov Transition Probabilities Characteristic, there is following property：On the non-directed graph V ∪ { Δ } of addition assisted tag, set and define q (Δ, Δ)=1 HeThen q (i, Δ) represents the probability that a random walk person leaves from figure G.
As illustrated, adding two kinds of assisted tag points in patent of the present invention, a kind of linked without label seed node, is claimed To eliminate node Δ, when random walk person reaches these nodes, it will link on the nodes, and corresponding probability can disappear Remove, that is to say, that a random walk person will not reach such node；Another kind is to link label node, referred to as retains nodeRandom walk person will be remained when reaching these nodes, these nodes are referred to as destination node.We also define SMFor Relative to VMReservation node, ciFor the probability left from V interior joints.Then random walk person is from node vi∈ V are in V ∪ { Δ } ∪ SM transition probability may be defined as：
Patent of the present invention passes through settingAs random walk person from node vi∈ V, which set out, reaches auxiliary reservation node's Probability, then have：
Wherein, to gauge method(N=| V | be the quantity of node) vector mark can be expressed as
Wherein,Be withThe N-dimensional row vector of composition.DcIt is with diagonal element ciPair of composition Angular moment battle array, i.e. Dc=diag (c1, c2..., cN), and I is N × N unit matrix.Transition probability matrix P=[pij]N×NIt is pixel Between weight matrix W row normalized matrix：
pij=wij/di
In patent of the present invention, we utilize and averagely reach probabilityThat is a random walk person is from node viSet out arrival Node label l is distributed in the reservation node that label is, expressionkPossibility probability.Then formula is changed into
Wherein E=I- (I-Dc)P。
Therefore in patent of the present invention, we provide this probability of stabilityCalculation formula it is as follows：
WhereinIt is to work asWhenWith in the case of otherColumn vector, ZkIt is that normalization is normal Number.Last each node vi=V distribution label result (i.e. the segmentation result of image) is determined by following formula：
Patent of the present invention is in terms of splitting with abundant texture information image, than existing random walk image Segmentation Technology It is more superior.The image segmentation of patent of the present invention is good, accuracy is high, has the characteristics of high robust and high stability.This hair Bright patent is to general pattern and the image with abundant texture information is all with good segmentation result.
It is not only obvious to general cutting object compared to other existing random walk image partition methods, patent of the present invention Image is accurately split, and the image to be split containing elongated complicated part also can accurately be split, and splits As a result accuracy is higher.
In image segmentation, accuracy is to weigh the standard of image segmentation accuracy, emphasizes segmentation figure picture and graphics standard point Segmentation difference between cutting.Accuracy is higher, represents the image after segmentation closer to Standard Segmentation.Accuracy in the present invention is determined Justice is：
Wherein StpIt is the percentage in the region correctly split；StnIt is the background percentage correctly split；SfpIt is erroneous segmentation Target area percentage；SfnIt is the background percentage of erroneous segmentation.Current Fig. 3 texture images are illustrated：
It can be calculated by above-mentioned formula：
The accuracy of the present invention is 98.83%, and the accuracy of traditional random walk method is 80.62%, specifically be see the table below：
Comparative approach Accuracy (%)
Traditional random walk method 80.62
The present invention 98.83
It is obvious to a person skilled in the art that the invention is not restricted to the details of foregoing example embodiment, Er Qie In the case of without departing substantially from spirit or essential attributes of the invention, the present invention can be realized in other specific forms.Therefore, no matter From the point of view of which point, embodiment all should be regarded as exemplary, and be nonrestrictive, the scope of the present invention is by appended power Profit requires rather than preceding description limits, it is intended that all in the implication and scope of the equivalency of claim by falling Change is included in the present invention.Any reference in claim should not be considered as to the involved claim of limitation.
Moreover, it will be appreciated that although the present specification is described in terms of embodiments, not each embodiment is only wrapped Containing an independent technical scheme, this narrating mode of specification is only that those skilled in the art should for clarity Using specification as an entirety, the technical solutions in the various embodiments may also be suitably combined, forms those skilled in the art It is appreciated that other embodiment.

Claims (6)

1. a kind of image partition method based on partial structurtes information around random walk combination pixel, specific algorithm flow bag Include following steps：
S1：Original image is inputted, program reads image to be split；
S2：Auxiliary is scribbled to add tag along sort and add assisted tag node and link tag along sort in image to be split Node, the tag along sort are divided into background label and prospect label；
S3：Using pixel information and structure tensor, composition novel local structural information defines similar between pixel in graph theory Property weighting function wij, to represent the similarity degree between them；
S4：Random walk person is set between pixel node in the image of addition assisted tag node with Markov property Transition probability；
S5：The transition probability between random walk person's pixel is solved using the Di Li Crays boundary function of combination；
S6：Using formula calculate random walk person from the node without label into the probability in all label nodes most Big value；To reach label node maximum probability as distribution tag along sort that principle is each pixel；
S7：It is recycled to untill whether there is label node all distributes to obtain tag along sort；To the contingency table of each pixel distribution Label complete final segmentation to image.
A kind of 2. image segmentation side based on partial structurtes information around random walk combination pixel according to claim 1 Method, it is characterised in that：The image that will split is mapped to a non-directed graph G=(V, E), by node v ∈ V and border Composition.Each node viRepresentative image pixel xi, eijRepresent two vertex vsiAnd vjLink, wij∈ W are represented between two summits Similarity degree, w hereij>=0, and wij≥wji。
A kind of 3. image segmentation side based on partial structurtes information around random walk combination pixel according to claim 2 Method, it is characterised in that：The degree of similarity represents that weighting function equation is as follows with dimensional Gaussian weight function：
Wherein IiAnd IlIt is the rgb value or gray value of the pixel of image, and JiAnd JjRepresent the structure tensor of partial structurtes information Scaling vector.
JiAnd JjRepresent the scaling vector of the structure tensor of partial structurtes information.In formula, | | Ji-Jj| | represent standard L2European model Number.
Wherein for two-dimensional image I：Ω→R2, arbitrarily certain gradient vector for putting pixel is in the pictureThen should Point tensor be：
And from det G=0, G only has a nonzero eigenvalue, therefore, tensor can only describe pixel one-dimentional structure and Direction, and the multidimensional information around pixel can not be described；
In order to embody partial structurtes information, filtering technique can be used to carry out matrix field data smooth, will it is smooth after filtering after Tensor be defined as structure tensor, its expression formula is：
Wherein hσFor the adaptive-filtering function of definition, its form is：
In formula, r is spatial frequency, and σ is variances sigma/(λ of gaussian kernel function1+λ2) size determine the effect model of filter function Enclose, if σ/(λ1+λ2) larger, illustrate that characteristic value sum is larger, partial structurtes are uniform domain, then using larger filter range； If conversely, σ/(λ1+λ2) smaller, it is edge or angle point region to illustrate partial structurtes, then using less filter range, in addition, ginseng Number θ represents the edge direction at (x, y) place, in the same direction with the characteristic vector corresponding to smaller characteristic value；
So, the structure tensor T based on new adaptive anisotropic filtering function can be written as：
In pretreatment stage, first using anisotropic filtering function mentioned above, image is transformed into structure tensor space, Then the mark and uniformity of the counter structure tensor of each pixel are asked for, obtains scaling vector J (x, y)；
Utilize the mark trG of structure tensorσ=λ1+λ2To describe the intensity of localized variation；Define coh Gσ=(λ1-λ2)2To portray office The uniformity of portion's structure；
Define J (x, y)=[trGσ coh Gσ], scaling vector J (x, y) illustrates the adjacent structural information of pixel, has more preferable Picture structure ability to express, the expression-form that can obtain connection weight between the pixel needed for Random Walk Algorithm is：
Wherein IiAnd IjIt is the rgb value or gray value of the pixel of image, and JiAnd JjRepresent the structure tensor of partial structurtes information Scaling vector.
Constant is arranged to ρ=10-7, control parameter is arranged toControl parameter σ1, σ2Taken by experiment Different values are contrasted, and actual result can be set, when in image split-run testWhen, segmentation effect It is most stable.
A kind of 4. image segmentation side based on partial structurtes information around random walk combination pixel according to claim 1 Method, it is characterised in that：The sub- Markov Transition Probabilities property definition：
Q is expressed as transition probability of no label node on V；By adding assisted tag point Δ and SM, according to Markov switching The characteristic of probability, there is following property：The V ∪ { Δ } on the non-directed graph of addition assisted tag, set and define q (Δ, Δ)=1 HeThen q (i, Δ) represents the probability that a random walk person leaves from figure G.
A kind of 5. image segmentation side based on partial structurtes information around random walk combination pixel according to claim 1 Method, it is characterised in that：The random walk person is from node vi∈ V are in V ∪ { Δ } ∪ SMTransition probability may be defined as：
As random walk person from node vi∈ V, which set out, reaches auxiliary reservation nodeProbability, then have：
Wherein, to gauge method(N=| V | be the quantity of node) vector mark can be expressed as
Wherein,Be withThe N-dimensional row vector of composition；DcIt is with diagonal element ciComposition to angular moment Battle array, i.e. Dc=diag (c1, c2..., cN), and I is N × N unit matrix；
Transition probability matrix P=[pij]N×NIt is the row normalized matrix of weight matrix W between pixel：
pij=wij/di
Utilize average arrival probabilityThat is a random walk person is from node viSet out and reach tape label lkReservation node, represent Distribute to node label lkPossibility probability；Then formula is changed into：
Wherein E=I- (I-Dc)P。
A kind of 6. image segmentation side based on partial structurtes information around random walk combination pixel according to claim 5 Method, it is characterised in that：The probability of stabilityCalculation formula it is as follows：
WhereinIt is to work asWhenWith in the case of otherColumn vector, ZkIt is normaliztion constant；
Last each node vi=V distribution label result (i.e. the segmentation result of image) is determined by following formula：
CN201711019238.9A 2017-10-27 2017-10-27 A kind of image partition method based on partial structurtes information around random walk combination pixel Pending CN107818566A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201711019238.9A CN107818566A (en) 2017-10-27 2017-10-27 A kind of image partition method based on partial structurtes information around random walk combination pixel

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201711019238.9A CN107818566A (en) 2017-10-27 2017-10-27 A kind of image partition method based on partial structurtes information around random walk combination pixel

Publications (1)

Publication Number Publication Date
CN107818566A true CN107818566A (en) 2018-03-20

Family

ID=61603358

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201711019238.9A Pending CN107818566A (en) 2017-10-27 2017-10-27 A kind of image partition method based on partial structurtes information around random walk combination pixel

Country Status (1)

Country Link
CN (1) CN107818566A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109685763A (en) * 2018-11-12 2019-04-26 浙江德尔达医疗科技有限公司 Three dimensional CT boniness block automatic division method, device, computer equipment and storage medium
CN112102343A (en) * 2020-08-12 2020-12-18 海南大学 Ultrasound image-based PTC diagnostic system

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170004627A1 (en) * 2013-05-22 2017-01-05 Siemens Aktiengesellschaft Parameter estimation for mesh segmentation using random walks
CN106780518A (en) * 2017-02-10 2017-05-31 苏州大学 A kind of MR image three-dimensional interactive segmentation methods of the movable contour model cut based on random walk and figure
CN107016682A (en) * 2017-04-11 2017-08-04 四川大学 A kind of notable object self-adapting division method of natural image

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170004627A1 (en) * 2013-05-22 2017-01-05 Siemens Aktiengesellschaft Parameter estimation for mesh segmentation using random walks
CN106780518A (en) * 2017-02-10 2017-05-31 苏州大学 A kind of MR image three-dimensional interactive segmentation methods of the movable contour model cut based on random walk and figure
CN107016682A (en) * 2017-04-11 2017-08-04 四川大学 A kind of notable object self-adapting division method of natural image

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
XINGPING DONG 等: ""Segmentation Using SubMarkov Random Walk"", 《EMMCVPR 2015: ENERGY MINIMIZATION METHODS IN COMPUTER VISION AND PATTERN RECOGNITION》 *
片兆宇: ""图像分割若干问题的研究与应用"", 《中国博士学位论文全文数据库 信息科技辑》 *
郭翠梅: ""基于随机游走的图像分割技术研究"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109685763A (en) * 2018-11-12 2019-04-26 浙江德尔达医疗科技有限公司 Three dimensional CT boniness block automatic division method, device, computer equipment and storage medium
CN112102343A (en) * 2020-08-12 2020-12-18 海南大学 Ultrasound image-based PTC diagnostic system
CN112102343B (en) * 2020-08-12 2024-03-29 海南大学 PTC diagnostic system based on ultrasonic image

Similar Documents

Publication Publication Date Title
CN106203430B (en) A kind of conspicuousness object detecting method based on foreground focused degree and background priori
CN104809187B (en) A kind of indoor scene semanteme marking method based on RGB D data
CN104537676B (en) Gradual image segmentation method based on online learning
CN105374033B (en) SAR image segmentation method based on ridge ripple deconvolution network and sparse classification
CN104504734B (en) A kind of color of image transmission method based on semanteme
CN106709568A (en) RGB-D image object detection and semantic segmentation method based on deep convolution network
CN108805957A (en) A kind of vector drawing generating method and system based on bitmap images adaptivenon-uniform sampling
CN105373777A (en) Face recognition method and device
CN109711268B (en) Face image screening method and device
CN111046868B (en) Target significance detection method based on matrix low-rank sparse decomposition
CN105787948A (en) Quick graph cutting method based on multiple deformation resolutions
CN106570874A (en) Image marking method combining local image constraint and overall target constraint
CN103593855A (en) Clustered image splitting method based on particle swarm optimization and spatial distance measurement
CN104408733A (en) Object random walk-based visual saliency detection method and system for remote sensing image
CN102663762B (en) The dividing method of symmetrical organ in medical image
CN108959379A (en) A kind of image of clothing search method of view-based access control model marking area and cartographical sketching
CN107909079A (en) One kind collaboration conspicuousness detection method
CN112700529A (en) Method and system for generating three-dimensional model according to standard document
CN107818566A (en) A kind of image partition method based on partial structurtes information around random walk combination pixel
Solomon et al. Continuous-flow graph transportation distances
CN104408731A (en) Region graph and statistic similarity coding-based SAR (synthetic aperture radar) image segmentation method
Zhang et al. Pointot: Interpretable geometry-inspired point cloud generative model via optimal transport
CN102542590B (en) High-resolution SAR (Synthetic Aperture Radar) image marking method based on supervised topic model
CN105894519A (en) Robustness image segmentation algorithm based on low rank recovery
CN102982567A (en) Plasmodium collision detection excluding method based on statistic analysis

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
WD01 Invention patent application deemed withdrawn after publication
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20180320