CN107330897A - Image partition method and its system - Google Patents
Image partition method and its system Download PDFInfo
- Publication number
- CN107330897A CN107330897A CN201710404301.4A CN201710404301A CN107330897A CN 107330897 A CN107330897 A CN 107330897A CN 201710404301 A CN201710404301 A CN 201710404301A CN 107330897 A CN107330897 A CN 107330897A
- Authority
- CN
- China
- Prior art keywords
- models
- energy functional
- pixel
- super
- matrix
- 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.)
- Granted
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
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of image partition method and its system, method includes：The salient region of target image is detected, the initialization boundary curve of target area is obtained；According to the energy functional number of LIF models and the energy functional number of DRLSE models, new energy functional number is generated；According to new the energy functional number and default iterations, the initialization boundary curve is developed, the boundary curve after being developed；Image segmentation is carried out according to the boundary curve after the evolution.By carrying out conspicuousness detection, initial curve is started from the adjacent edges of target area, greatly save the time of evolution, improve the degree of accuracy of segmentation；The Level Set Method combined by local message and gradient information, is capable of the image of effectively Ground Split background information complexity and weak boundary.
Description
Technical field
The present invention relates to technical field of image processing, more particularly to a kind of image partition method and its system.
Background technology
In the research and application of image, people are often interested in some of image part, these parts interested
The specific, region (single area can be corresponded to, multiple regions can also be corresponded to) with special nature in image is generally corresponded to,
Referred to as target or prospect；And other parts are referred to as the background of image.In order to recognize and analyze target, it is necessary to target from a width
Isolated out in image, here it is the problem of image segmentation will be studied.Image segmentation be exactly divide the image into that several are specific,
Region with unique properties and the technology and process for proposing interesting target.Image segmentation is image recognition and computer vision
Vital pretreatment.Correct identification has been impossible to without correctly segmentation.But, carrying out the only foundation of segmentation is
The brightness of pixel and color in image, when automatically processing segmentation by computer, it will run into all difficulties.For example, uneven illumination
There is unsharp part, and shade etc. in even, noise influence, image, it occur frequently that segmentation errors.Therefore image segmentation
It is the technology for needing further to study.It is desirable to introduce the method for some artificial knowledge guiding and artificial intelligence, for entangling
Mistake in just some segmentations, is up-and-coming method, but which again increases solve the complex nature of the problem.Image is split
Image understanding and the premise of identification.It is always image procossing and computer vision field as the basic link of image procossing
Focus and difficulties.The movable contour model realized using Level Set Method is paid close attention to by numerous scholars in recent years.
Image segmentation has three kinds with different approach：
One is incorporating each pixel into respective objects or the Pixel Clustering in region, i.e. field method；
The second is realizing the boundary method of segmentation by directly determining the border in region；
The third is detection of edge pixels, then edge pixel is connected into the formation segmentation of composition border first.
The Level Set Method that Osher and Sethian are proposed first：It is based primarily upon curve evolvement and Level Set Theory.Water
The basic thought of flat collection is that the evolution curve or curved surface of image are embedded into higher one-dimensional level set, high dimension curve or
The evolution curve of curved surface converts the partial differential equation of higher-dimension, and final evolution curve is obtained by solving partial differential equation.
D.Mumford et al. proposes a kind of geometric active contour model (Mumford-Shah, MS), and the model is based on fracture mechanics
Variation energy equation, can quickly be broken in evolutionary process, merge, but the energy term of MS models is difficult to find simply
Numerical radius, hinders MS model developments.
Caselles et al. proposes the Image Segmentation Model based on Level Set Method --- geodetic skeleton pattern (Geodesic
Active Contour, GAC), the model allows energy functional carry out change in topology on the basis of conventional model, still
Otherwise the initialization curve of the model can not must naturally carry out topological structure completely in the inside or outside of target to be cut
Change.
2001, Chan et al. proposed a kind of C-V models, and this model is added in energy function comprising length of curve
With the penalty term of local region area, the model is to relatively good containing noisy image effect, but for the uneven figure of gray scale
Picture, the C-V models of the global information based on image, has tended not to good segmentation effect.
Repeat to initialize to overcome traditional level set movements to need, Li et al., which proposes one kind, to be reinitialized
Regular level set movements (DRLSE) model of distance, its main thought is to add internal energy in energy function equation to punish
Penalize item, internal energy penalty term cause level set curve in evolutionary process from symbolic measurement will not deviation it is too remote, all the time
Remain symbolic measurement or approximation sign distance function so that evolution curve need not be reinitialized.Although this
Method, which is avoided, to be reinitialized, but when splitting the image that background information is complicated or gray scale is uneven, segmentation can be caused bent
Line deviates target area, so as to the segmentation of mistake occur.
Conspicuousness detection is one, current computer field study hotspot, for selecting the related content in visual scene
(thing or region) is used as human eye vision notice region.Salient region detection can as image pretreatment stage, exist
More and more important effect has been given play in terms of image retrieval, classification, the editor of image and image segmentation.Conspicuousness model can
Be divided into it is top-down and from bottom to top both, top-to-bottom method is to obtain figure with the Advanced information retrieval of image
The saliency value of picture, bottom-to-top method uses the low-level information such as color of image, and distance etc. obtains the saliency value of image.It is existing
Some marking area detection algorithms by calculate each image region with its a range of adjacent area contrast come
The conspicuousness of the image region is measured, salient region detecting method plays following field the great effect of benefiting：Image
Image scaling that segmentation, object detection, content are kept etc..
The mankind can rapidly and accurately recognize the marking area in field of vision.This ability pair of the mankind is simulated on machine
It is vital in enabling the machine to handle vision content as the mankind.It is interior in the past few decades, it is existing substantial amounts of aobvious
Work property detection method, which is entered, to be published.Major part in these methods is all intended to predict human eye vision blinkpunkt forefathers
Propose many methods, for example, Itti et al. propose based on visual attention model from bottom to top, to the color of image,
Direction etc. is extracted, and obtains the single notable figure of image.The quaternary Fourier transformation by calculating image that Guo et al. is proposed
Phase spectrum, obtain the notable figure of image, each four-tuple includes color, brightness and Vector Groups.
The salient region detection algorithm mode classification that Borji et al. was proposed in 2014, this model is by drawing
The profile of object, detects obvious object.It can be seen that model from bottom to top is this several years trend, top-down model is profit
There is no priori to utilize with the attention model of vision, available information is seldom, so this model is too complicated.
The content of the invention
The technical problems to be solved by the invention are：There is provided a kind of image partition method and its system, it is possible to resolve existing skill
The problem of image complicated to background information and weak boundary splits inaccurate in art.
In order to solve the above-mentioned technical problem, the technical solution adopted by the present invention is：A kind of image partition method, including：
The salient region of target image is detected, the initialization boundary curve of target area is obtained；
According to the energy functional number of LIF models and the energy functional number of DRLSE models, new energy functional number is generated；
According to new the energy functional number and default iterations, the initialization boundary curve is developed,
Boundary curve after being developed；
Image segmentation is carried out according to the boundary curve after the evolution.
The invention further relates to a kind of image segmentation system, including：
Detection module, the salient region for detecting target image obtains the initialization boundary curve of target area；
Generation module, for the energy functional number and the energy functional number of DRLSE models according to LIF models, is generated newly
Energy functional number；
Genetic module, for according to new the energy functional number and default iterations, to the initialization border
Curve is developed, the boundary curve after being developed；
Split module, for carrying out image segmentation according to the boundary curve after the evolution.
The beneficial effects of the present invention are：By carrying out conspicuousness detection, it is easy to obtain the marginal information of image, also
It is the gradient information of target area, the information of background area can be excluded well, initial curve is started from target area
Adjacent edges, greatly save the time of evolution, improve the degree of accuracy of segmentation, final curves is located at target area well
Domain；The Level Set Method combined by local message and gradient information, ensure that segmentation precision well, being capable of effective Ground Split
The image of background information complexity and weak boundary；The present invention can improve the efficiency and accuracy of image segmentation, and no matter in time,
Or significantly it is better than DRLSE models on parted pattern, for polytype picture, with higher computational efficiency and accurately
Property.
Brief description of the drawings
Fig. 1 is a kind of flow chart of image partition method of the embodiment of the present invention one；
Fig. 2 is a kind of structural representation of image segmentation system of the invention；
Fig. 3 is the system structure diagram of the embodiment of the present invention three.
Label declaration：
1st, detection module；2nd, generation module；3rd, genetic module；4th, module is split；
101st, the first division unit；102nd, the second division unit；103rd, the first construction unit；104th, the second construction unit；
105th, the first computing unit；106th, the 3rd construction unit；107th, normalization unit；
108th, the 4th construction unit；109th, unit is optimized；110th, evolution updating block；111st, first unit is obtained；
201st, generation unit；202nd, second unit is obtained；203rd, the 3rd unit is obtained；204th, the 4th unit is obtained.
Embodiment
To describe technology contents, the objects and the effects of the present invention in detail, below in conjunction with embodiment and coordinate attached
Figure is explained in detail.
The design of most critical of the present invention is：The detection of salient region is carried out by cellular automata；Pass through local letter
The image of the Level Set Method that breath and gradient information are combined, effective Ground Split background information complexity and weak boundary.
Explanation of nouns：
Cellular automata：Cellular automata is made up of grid, there is discrete state on grid, according to corresponding rule
Then, cellular can update the state of oneself in the discrete time.The current state of each cellular by previous moment state
And the Determines of the previous moment of cellular adjacent thereto.
Super-pixel：By a series of cell that positions are adjacent and the similar pixel of color, brightness, Texture eigenvalue is constituted
Domain.
Referring to Fig. 1, a kind of image partition method, including：
The salient region of target image is detected, the initialization boundary curve of target area is obtained；
According to the energy functional number of LIF models and the energy functional number of DRLSE models, new energy functional number is generated；
According to new the energy functional number and default iterations, the initialization boundary curve is developed,
Boundary curve after being developed；
Image segmentation is carried out according to the boundary curve after the evolution.
It was found from foregoing description, the beneficial effects of the present invention are：The time of evolution can be greatlyd save, segmentation is improved
The degree of accuracy；Can effectively Ground Split background information be complicated and image of weak boundary.
Further, local message LIF methods are added on the basis of DRLSE methods, LIF methods are well details
Information is as energy term, and DRLSE methods are added to gradient information in energy equation as penalty term, not only control evolution
Speed, direction also rests on evolution curve on target area boundaries well.
Further, it is described " according to cellular automata, to detect the salient region of target image, obtain target area
Initialize boundary curve " be specially：
Target image is divided into N number of super-pixel；
According to K mean algorithms, the super-pixel positioned at target image edge is divided into K classes；
According to the K class super-pixel after division, K global color disparity map and its corresponding notable value matrix are built；
The space length between super-pixel in each super-pixel and K class super-pixel, builds weight matrix；
According to the notable value matrix and weight matrix, calculating obtains notable figure；
According to the neighbouring relations and space length between super-pixel, define influence of the super-pixel to another super-pixel because
Son, and build factor of influence matrix；
According to default degree matrix, the factor of influence matrix is normalized, the factor of influence after being normalized
Matrix；
According to the factor of influence in factor of influence matrix, the confidence level of each super-pixel current state is calculated, and builds confidence
Spend matrix；
The confidence level matrix is optimized, the confidence level matrix after being optimized；
According to the confidence level matrix after the factor of influence matrix after normalization and optimization, the notable figure is developed more
Newly, the final notable figure of target image is obtained；
According to the final notable figure, the initialization boundary curve of target area is obtained.
Seen from the above description, the notable figure that target image is rapidly and accurately found out by the use of cellular automata is used as new model
Initial evolution curve, it is to avoid DRLSE models choose initial segmentation curve this drawback manually, greatly save the time, and
Segmentation precision is high.
Further, it is described " according to the energy functional number of LIF models and the energy functional number of DRLSE models, to generate newly
Energy functional number " is specially：
According to the energy functional number of LIF models and the energy functional number of DRLSE models, new energy functional number, institute are generated
Stating new energy functional number isWherein, ELIFFor the energy functional number of LIF models, EDRLSEFor
The energy functional number of DRLSE models, η is the weight of default correspondence LIF models, and ρ is the power of default correspondence DRLSE models
Weight；
According to the energy functional number of LIF models, the level set movements equation of LIF models is obtained；
According to the energy functional number of DRLSE models, the level set movements equation of DRLSE models is obtained；
According to the level set movements of the new energy functional number, the level set movements equation of LIF models and DRLSE models
Equation, obtains new level set movements equation.
Seen from the above description, local message LIF methods are added on the basis of DRLSE methods, LIF methods are good
Using detailed information as energy term, DRLSE methods are added to gradient information in energy equation as penalty term, are not only controlled
The speed of evolution, direction also rests on evolution curve on target area boundaries well.
Further, it is described " according to new the energy functional number and default iterations, to the initialization border
Curve is developed, the boundary curve after being developed " be specially：
According to new the level set movements equation and default iterations, the initialization boundary curve is drilled
Change, the boundary curve after being developed.
Seen from the above description, the new energy function combined using LIF models and DRLSE models ensure that point well
The precision cut.
Fig. 2 is refer to, the invention also provides a kind of image segmentation system, including：
Detection module, the salient region for detecting target image obtains the initialization boundary curve of target area；
Generation module 2, for the energy functional number and the energy functional number of DRLSE models according to LIF models, is generated newly
Energy functional number；
Genetic module, for according to new the energy functional number and default iterations, to the initialization border
Curve is developed, the boundary curve after being developed；
Split module, for carrying out image segmentation according to the boundary curve after the evolution.
Further, the detection module is specifically for according to cellular automata, detecting the salient region of target image,
Obtain the initialization boundary curve of target area.
Further, the detection module includes：
First division unit, for target image to be divided into N number of super-pixel；
Second division unit, for according to K mean algorithms, the super-pixel positioned at target image edge to be divided into K classes；
First construction unit, for according to the K class super-pixel after division, building K global color disparity map and its correspondingly
Notable value matrix；
Second construction unit, for the space length between the super-pixel in each super-pixel and K class super-pixel, structure
Build weight matrix；
First computing unit, for according to the notable value matrix and weight matrix, calculating to obtain notable figure；
3rd construction unit, for according to the neighbouring relations and space length between super-pixel, defining a super-pixel to another
The factor of influence of one super-pixel, and build factor of influence matrix；
Normalization unit, for according to default degree matrix, being normalized to the factor of influence matrix, obtaining normalizing
Factor of influence matrix after change；
4th construction unit, for the factor of influence in factor of influence matrix, calculates each super-pixel current state
Confidence level, and build confidence level matrix；
Optimize unit, for being optimized to the confidence level matrix, the confidence level matrix after being optimized；
Evolution updating block, for according to the confidence level matrix after the factor of influence matrix after normalization and optimization, to institute
State notable figure and carry out evolution renewal, obtain the final notable figure of target image；
First obtains unit, for according to the final notable figure, obtaining the initialization boundary curve of target area.
Further, the generation module includes：
Generation unit, for the energy functional number and the energy functional number of DRLSE models according to LIF models, is generated newly
Energy functional number, the new energy functional number isWherein, ELIFEnergy for LIF models is general
Function, EDRLSEFor the energy functional number of DRLSE models, η is the weight of default correspondence LIF models, and ρ is default correspondence
The weight of DRLSE models；
Second obtains unit, for the energy functional number according to LIF models, obtains the level set movements equation of LIF models；
3rd obtains unit, for the energy functional number according to DRLSE models, obtains the level set movements of DRLSE models
Equation；
4th obtains unit, for according to the new energy functional number, the level set movements equation of LIF models and
The level set movements equation of DRLSE models, obtains new level set movements equation.
Further, the genetic module is specifically for according to the new level set movements equation and default iteration time
Number, develops to the initialization boundary curve, the boundary curve after being developed.
Embodiment one
Fig. 1 is refer to, embodiments of the invention one are：A kind of image partition method, methods described is based on salient region
Detection and level set, comprise the following steps：
S1：The salient region of target image is detected, the initialization boundary curve of target area is obtained；In the present embodiment
Target area is the salient region in target image, and target area is also target to be split.
S2：According to the energy function of LIF models and the energy function of DRLSE models, new energy function is generated；
S3：According to the new energy function and default iterations, the initialization boundary curve is developed,
Boundary curve after being developed；
S4：Image segmentation is carried out according to the boundary curve after the evolution.
In step S1, salient region is the pixel of most attractive visual attention in picture, the mark of conspicuousness detection
It is accurate as follows：
A, protrusion object the most significant；
B, consistent highlighted whole obvious object；
C, the accurate border for meeting object；
D, higher noise immunity；
E, full resolution.
The present embodiment detects the salient region for obtaining target image using the conspicuousness detection algorithm of cellular automata.
Specifically, step S1 comprises the following steps：
S101：According to simple linear Iterative Clustering, target image is divided into N number of super-pixel；Simple linear iteration
Cluster that (simple linear iterative clustering, SLIC) algorithm structure is simple, it is necessary to parameter is few, can be effective
Ground divides the image into block of pixels with different size and shape.The side of object in figure has largely been pressed close on the border of super-pixel
Boundary, each super-pixel is a representative region, its not only color containing bottom, directional information, but also wrap
The structural information in middle level is contained.Final conspicuousness result of calculation can be ensured to thing using super-pixel as basic calculating unit
The boundary representation of body is more accurate.Further, the size of each super-pixel is 9*9 pixels in the present embodiment.Each super picture
Element represents a cellular.
S102：According to K mean algorithms, the super-pixel positioned at target image edge is divided into K classes；Preferably, K=3,
The super-pixel that target image edge will be located at is divided into three classes.
S103：According to the K class super-pixel after division, K global color disparity map (Global Color is built
Distinction, GCD) and its corresponding notable value matrix M, M=[mk,i]K×N, mk,iRepresent in k-th of global color disparity map
Middle super-pixel i significance value, can calculate according to the first formula and obtain；
First formula：
Wherein, pkFor the super-pixel sum of kth class super-pixel, k=1,2 ..., K, ‖ di, dj ‖ are super-pixel i and super-pixel
Euclidean distances of the j in CILELAB color spaces, i=1,2 ..., N, σ1It is default constant with β, in the present embodiment, σ1=
0.2, β=10.
S104：The space length between super-pixel in each super-pixel and K class super-pixel, builds weight matrix W.
K significantly value matrix M have been obtained in step s 103；The obtained GCD figures of edge cluster are based only on, as a result not
Can be satisfactory, but have the super-pixel of some pin-point accuracys in the every width figure of result, because the super-pixel after optimization has very
Big similitude, in order to optimize the GCD figures obtained based on edge, builds weight matrix W=[wk,i]K×NTo weigh different GCD figures
Between importance.Because the super-pixel point of image border is connected with each other, therefore by the super-pixel of image border as the back of the body
Scape seed；wk,iThe space length between super-pixel i and kth class background seed, i.e. super-pixel in kth class super-pixel is represented, can
Calculated and obtained according to the second formula；
Second formula：
Wherein, riAnd rjRespectively (super-pixel contains multiple pixels to super-pixel i and super-pixel j coordinate, with every
The average value of all pixels point of individual super-pixel as the super-pixel coordinate points), ‖ ri, rj ‖ is super-pixel i and super-pixel j
Euclidean distance, σ2For the constant of default control weight, in the present embodiment, σ2=1.3.
S105：According to notable the value matrix M and weight matrix W, calculating obtains notable figure Mbg, Mbg=[M1 bg,...,
MN bg]T, Mi bgCalculated and obtained according to the 3rd formula.
3rd formula：
GCD figures are constrained with space length, local area contrast can be strengthened, so as to improve the accurate of significance value
Property.By effectively utilizing the advantage of different GCD figures, the notable figure obtained based on background is more convincing.
S106：According to the neighbouring relations and space length between super-pixel, shadow of the super-pixel to another super-pixel is defined
Ring factor fi,j, and build factor of influence matrix F.
Neighbours' cellular of a cellular is defined, including the cellular adjacent with it and the cellular adjacent with it have together
The cellular at one edge.Simultaneously it is considered that the super-pixel of image border is connected with each other, therefore they are all taken as background kind
Son.Neighbours' cellular is not changeless to the influence power of cellular.Intuitively think if neighbours' cellular is to some cellular
There is more like color characteristic, bigger influence, the phase of any pair of cellular will be produced to the state of the cellular subsequent time
Like property weighed by the distance defined in CIELAB color spaces.
Therefore, by defining factor of influence fs of the super-pixel i to super-pixel ji,j, come build factor of influence matrix F=
[fi,j]N×N, fi,jIt can be calculated and obtained according to the 4th formula；
4th formula：
fi,j=0, otherwise
Wherein, NB (i) is the super-pixel i (set of cellular i) neighbours' super-pixel (neighbours' cellular), that is to say, that when super
Pixel j is not super-pixel i neighbours' super-pixel, then super-pixel i is 0 to super-pixel j factor of influence；‖ di, dj ‖ are super-pixel i
With Euclidean distances of the super-pixel j in CILELAB color spaces, σ3For the parameter of default control similarity measurement, this implementation
In example, σ3=0.1.F size and M or W size are irrelevant.
S107：According to default degree matrix D, the factor of influence matrix F is normalized, the shadow after being normalized
Ring factor matrix F*.Definition degree matrix D=diag { d1,d2,…dN, wherein, di=∑jfij；According to the 5th formula to the shadow
Factor matrix F is rung to be normalized；
5th formula：F*=D-1F
S108：According to the factor of influence in factor of influence matrix, the confidence level of each super-pixel current state is calculated, and is built
Confidence level Matrix C.
Because the state at each cellular lower a moment is together decided on by the state of current state and previous moment, it is therefore desirable to
Balance the two deciding factors.In color space, super-pixel and its neighbours' super-pixel have very big difference, under it
The main Determines by current time of the state at one moment, then it is likely to be assimilated by local environment, therefore sets up one
Individual confidence level Matrix C=diag { c1,c2,…cNThe renewal for advantageously promoting all cellulars is gone to develop, each cellular is current to it
The confidence level c of moment stateiIt can be calculated and obtained according to the 6th formula；
6th formula：
S109：The confidence level Matrix C is optimized, the confidence level Matrix C * after being optimized.In order to ensure ci
In default interval [b, a+b], according to the 7th formula to ciOptimize, obtain ci*, so that confidence level square after being optimized
Battle array C*=diag { c1*,c2*,…cN*}；
7th formula：
Wherein, j=1,2 ..., N, a and b are default constant, in the present embodiment, a=0.6, b=0.2.
Using the confidence level Matrix C * after optimization, cellular will automatically update next more accurate and stabilization state.
S110：According to the confidence level Matrix C * after the factor of influence matrix F * after normalization and optimization, to the notable figure
MbgEvolution renewal is carried out, the final notable figure of target image is obtained；
Factor of influence matrix F * after normalization is to weigh super-pixel by neighbours' super-pixel influence degree, the confidence after optimization
Degree Matrix C * is influence of the cellular to the cellular of later moment in time of previous moment, using F* and C* as weighing neighbours' cellular and previous
The cellular disturbance degree of the cellular state at moment each to current time.
Specifically, it is updated according to the 8th formula；
8th formula：Mt+1=C*Mt+(I-C*)F*Mt
Wherein, MtRepresent the more new state at current time, Mt+1The more new state of subsequent time is represented, is just as t=0
Beginning state, i.e. M0=Mbg, determined by the characteristic of image in itself, by default Nl=t+1 time step (NlBy super picture
The number N of element is determined), obtain final notable figure, namely salient region.
S111：According to the final notable figure, the initialization boundary curve of target area is obtained.
Based on the substantive characteristics of most of images, belonging to the super-pixel of prospect generally has similar color characteristic, utilizes neighbour
The intrinsic contact of super-pixel is occupied, individual layer cellular automata can strengthen the uniformity of the saliency value of similar area, and form one
Stable local environment.Secondly, conspicuousness target and its surrounding environment have very big otherness in color space, by similar
An obvious line of demarcation, can naturally occur between target and background in influence between neighbours' super-pixel.And cellular is automatic
Machine can be very good reinforcement prospect and suppress background, therefore, based on cellular automata, and intuitively update mechanism is designed to pass through
Interaction with neighbours utilizes the inherence of prominent object to be connected.It is this based on context propagate can by it is any it is given at first
The result entered is improved to the similar level with higher precision, and with higher accuracy and memory degree.
Because DRLSE models are manual selection initial segmentation curves, so initial segmentation curve must be in target to be split
Outside or inner side, cause segmentation time increase.And conspicuousness model of the present embodiment based on principle of cellular automation
The profile of target to be split substantially is found, is then not only greatlyd save as initial segmentation curve with salient region profile
Time, also improve the efficiency of algorithm.
For step S2, topography's fitting (LIF) model is the local gray level information structuring energy function using image.
It extracts the local message of image using gaussian kernel function, is put down with two local auto-adaptive approximation to function contoured interiors with outside
Equal gray value, minimizes energy function and obtains segmentation result.Assuming that target image is I：Ω→Οl(Ο represents the field of image,
L represents image I (x) dimension).Local fit I is constructed first at point xLIF, then by domain of definition Ω to x at
Local fit energy term ILIFIntegration is taken, the energy functional number of LIF models is obtained, as shown in the 9th formula；Topography is fitted energy
Formula is measured as shown in the tenth formula, the m in the tenth formula1、m2Calculated and obtained according to the 11st formula；
9th formula：
Tenth formula：ILIF=m1Hω(φ)+m2(1-Hω(φ))
11st formula：
Wherein, * represents convolution algorithm；Hω(φ) is Heaviside functions；m1(x), m2(x) it is zero level collection curvilinear inner
With the average of outer partial rectangular area, the two local rectangular portions are respectively With, wherein, WK(x) it is a rectangular window, is typically selected to have standard deviation sigma and size is
(4h+1) × (4h+1's) blocks Gauss window Rσ, wherein h is less than σ maximum integer.
The energy functional number of the LIF models is minimized, the level set movements equation of LIF models is obtained, the such as the 12nd is public
Shown in formula.
12nd formula：
LIF models are obvious to the image segmentation of certain kinds, such as the effect ratio split to the image of weak boundary
Preferably, but its have the shortcomings that segmentation result depend on initial profile size, shape and position, be easily trapped into local minimum
Problem, so the unreasonable of initial profile design can cause the result of mistake.
Movable contour model (active contour model) basic thought is that target side is expressed using full curve
Edge, and define an energy functional cause its independent variable include boundary curve, therefore cutting procedure be just changed into solution energy it is general
The process of the minimum value of letter, can typically be realized, energy by the corresponding Euler of solved function (Euler.Lagrange) equation
It is exactly the profile place of target to reach curve location during minimum.Level set is exactly a kind of typical movable contour model.But
Traditional Level Set Models need initialization operation repeatedly, add the time of evolution.
Therefore, the present embodiment also introduces level set movements (the Distance Regularized apart from regularization
Level Set Evolution Mode, DRLSE) model.
The principle of Level Set Models is that initial curve is constantly close to edge by iteration, is finally divided the image into out
Come.In parameter active contour model, the evolution of curve is realized by minimizing energy function finally to be split, and level
During diversity method is then 0 level set being embedded into curve in the curved surface of one 3-dimensional, by the curved surface in evolution 3-dimensional space, then from
0 level set is obtained in curved surface after evolution, this 0 level set is exactly the curve after developing.By in level set function profile die
A penalty term is added in type, it is to avoid the distance function of level set function principle tape symbol in evolutionary process, while also one
Item data so that evolution curve is developed towards objective contour direction of curve.This method very effectively avoids level
The problem of set function needs to reinitialize in evolutionary process.And the level set function of initialization is not limited to tape symbol
Distance function, used alternative manner also causes computational efficiency to be improved.But, one is exactly to scheme the problem of very serious
As that can have the uneven phenomenon of gray scale.In order to overcome the initialization to curve, use need not repeat the variation water of initialization
Flat collection model, the model overcomes the problem of diffusivity tends to infinity, adds penalty term, and correct for level set function with
Deviation between symbolic distance, improves the Evolution Rates of curve.Its energy functional number such as the 13rd formula apart from regularization
It is shown；
13rd formula：EDRLSE=μ EP+Eext
Wherein, EPIt is range calibrationization, it is to ensure that evolution curve is remained or close to symbol to add this
Distance function, its expression formula is as shown in the 14th formula；μ is default constant, for weighing apart from normalized importance；
EextIt is external energy function, is to ensure that evolution curve can be parked in target area well.
14th formula：
Wherein, "=" in the 14th formula isDefinition is represented, i.e., to EPIt is defined.Due to symbolic measurement
It must is fulfilled for | ▽ φ=1, therefore, the energy functional number of DRLSE models is as shown in the 15th formula；
15th formula：EDRLSE=μ EP(φ)+λL(φ)+αA(φ)
Wherein, λ>0, α is default real number, and the two parameters determine the proportion shared by L (φ) and A (φ), L (φ) and A
The expression formula of (φ) is as shown in the 16th formula；
16th formula：
Wherein, "=" in the 16th formula is alsoL (φ) and A (φ) are defined；δ (x) is single argument
Impulse function, can be obtained according to the 17th formula；Hω(x) it is Heaviside functions, can be obtained according to the 18th formula；
17th formula：
δ (x)=0, | x | ＞ ω
18th formula：
Hω(x)=1, x ＞ ω
Hω(x)=0, x ＜-ω
Wherein, ω is default constant；G is edge indicator function, definitionWherein, R is a mark
The gauss of distribution function that quasi- difference is σ, σ is scale parameter.G is the attribute function on image, and L (φ) makes curve towards mesh
Mark region is developed, and curve is parked at target.A (φ) determines the Evolution Rates of curve.
Therefore, with reference to the 16th formula, the 17th formula, the 18th formula, using most fast decrease speed process as gradient current
The level set movements equation of final evolution curve, i.e. DRLSE models is obtained, as shown in the 19th formula and the 20th formula；
19th formula：
20th formula：
Wherein, Δ represents Laplace operator.This model biggest advantage is to overcome traditional level set needs
The problem of reinitializing, makes curve in evolutionary process close to symbolic measurement, because evolution principle is to be based on gradient
Information, there is good localized effect to image border, using the partial gradient information of image without considering image
Global information, using this model to soft edge, the few image segmentation of local message is not very good, convergence rate
It is slow.So with reference to the local message of image, not merely only considering the gradient information of image, obtained parted pattern has very
Good segmentation effect.
Based on above-mentioned analysis, with reference to local message and gradient information, LIF models and DRLSE factors of a model are introduced, generation is new
Energy functional number, as shown in the 21st formula；
21st formula：
Wherein, ELIFFor the energy functional number of LIF models, EDRLSEFor the energy functional number of DRLSE models, η is default right
The weight of LIF models is answered, ρ is the weight of default correspondence DRLSE models.
According to the 9th formula, the 12nd formula, the 15th formula, the 19th formula and the 21st formula, it can obtain most
Whole evolution curvilinear equation, as shown in the 22nd formula；
22nd formula：
According to above-mentioned formula, by new level set movements equation (the 22nd formula) it is discrete be finite difference equations, such as
Shown in 23rd formula；
23rd formula：
Wherein,Approximate solution on the right of the new level set movements equation (the 22nd formula) of expression, when Δ t is represented
Between step-length,The evolution curve at current time is represented,Represent the evolution curve of subsequent time.As can be seen that increase step-length
The Evolution Rates of curve can be accelerated, the Evolution Rates of curve can be slowed down by reducing step-length.
For step S3 and step S4, the initialization boundary curve obtained with step S1, i.e. conspicuousness detection contour curve
As initial profile curve, the evolution that new EVOLUTION EQUATION instructs curve is obtained with step S2.Finally according to the curve after evolution
Carry out image segmentation
The present embodiment finds the profile of target to be split substantially with the conspicuousness model based on principle of cellular automation, then
With salient region profile as initial segmentation curve, the time is not only greatlyd save, the efficiency of algorithm is also improved.
Local message LIF methods are added on the basis of DRLSE methods, LIF methods well using detailed information as energy term,
DRLSE methods are added to gradient information in energy equation as penalty term, not only control the speed of evolution, direction, also very
Evolution curve is rested on target area boundaries well.The Level Set Method combined by local message and gradient information, very
Segmentation precision is ensure that well, can effectively Ground Split background information be complicated and image of weak boundary.
Embodiment two
The present embodiment is a concrete application scene of above-described embodiment.
First, the parameter of the new level set movements equation of setting, η=0.1, ρ=0.9, time step Δ t=1, μ=
0.2, λ=5, α=1.5, iterations is 11.The notable figure M of target image is obtained according to cellular automatat+1, i.e. conspicuousness area
Domain, obtains notable figure Mt+1Average Mmean, using the average as threshold value, target image is divided into by two parts according to threshold value, will
The curve of division initializes boundary curve as initial profile curve.According to the 11st formula and the 16th formula, count respectively
Calculate m1And m2, L (φ) and A (φ), then according to new level set movements equation and its finite difference equations, every Δ t=1s
Develop a level set function.If the number of times developed is unsatisfactory for iterations, continue evolution curve, until meeting iteration time
Number, obtains final evolution curve, that is, final segmentation curve, and image segmentation is carried out according to segmentation curve.
Embodiment three
Fig. 3 is refer to, the present embodiment is a kind of image segmentation system of correspondence above-described embodiment, including：
Detection module 1, the salient region for detecting target image obtains the initialization boundary curve of target area；
Generation module 2, for the energy functional number and the energy functional number of DRLSE models according to LIF models, is generated newly
Energy functional number；
Genetic module 3, for according to new the energy functional number and default iterations, to the initialization border
Curve is developed, the boundary curve after being developed；
Split module 4, for carrying out image segmentation according to the boundary curve after the evolution.
Further, the detection module 1 is specifically for according to cellular automata, detecting the conspicuousness area of target image
Domain, obtains the initialization boundary curve of target area.
Further, the detection module 1 includes：
First division unit 101, for target image to be divided into N number of super-pixel；
Second division unit 102, for according to K mean algorithms, the super-pixel positioned at target image edge to be divided into K
Class；
First construction unit 103, for according to the K class super-pixel after division, building K global color disparity map and its right
The notable value matrix answered；
Second construction unit 104, for the space between the super-pixel in each super-pixel and K class super-pixel away from
From structure weight matrix；
First computing unit 105, for according to the notable value matrix and weight matrix, calculating to obtain notable figure；
3rd construction unit 106, for according to the neighbouring relations and space length between super-pixel, defining a super-pixel pair
The factor of influence of another super-pixel, and build factor of influence matrix；
Normalization unit 107, for according to default degree matrix, being normalized, obtaining to the factor of influence matrix
Factor of influence matrix after normalization；
4th construction unit 108, for the factor of influence in factor of influence matrix, calculates each super-pixel current state
Confidence level, and build confidence level matrix；
Optimize unit 109, for being optimized to the confidence level matrix, the confidence level matrix after being optimized；
Evolution updating block 110, it is right for according to the confidence level matrix after the factor of influence matrix after normalization and optimization
The notable figure carries out evolution renewal, obtains the final notable figure of target image；
First obtains unit 111, for according to the final notable figure, obtaining the initialization boundary curve of target area.
Further, the generation module 2 includes：
Generation unit 201, for the energy functional number and the energy functional number of DRLSE models according to LIF models, generation is new
Energy functional number, the new energy functional number isWherein, ELIFFor the energy of LIF models
Functional, EDRLSEFor the energy functional number of DRLSE models, η is the weight of default correspondence LIF models, and ρ is default correspondence
The weight of DRLSE models；
Second obtains unit 202, for the energy functional number according to LIF models, obtains the level set movements side of LIF models
Journey；
3rd obtains unit 203, and for the energy functional number according to DRLSE models, the level set for obtaining DRLSE models is drilled
Change equation；
4th obtains unit 204, for according to the new energy functional number, the level set movements equation of LIF models and
The level set movements equation of DRLSE models, obtains new level set movements equation.
Further, the genetic module 3 is specifically for according to the new level set movements equation and default iteration
Number of times, develops to the initialization boundary curve, the boundary curve after being developed.
In summary, the present invention is provided a kind of image partition method and its system, by carrying out conspicuousness detection, hold very much
Easily obtain image marginal information, that is, target area gradient information, the information of background area can be excluded well,
Initial curve is started from the adjacent edges of target area, greatly save the time of evolution, improve the degree of accuracy of segmentation, make most
Finale line is located at target area well；Target to be split is found substantially by the conspicuousness model based on principle of cellular automation
Profile, then not only greatlyd save the time as initial segmentation curve with salient region profile, also improve algorithm
Efficiency；Local message LIF methods are added on the basis of DRLSE methods, LIF methods assign detailed information as energy well
, DRLSE methods are added to gradient information in energy equation as penalty term, not only control the speed of evolution, direction,
Evolution curve is rested on target area boundaries well；The Level Set Method combined by local message and gradient information,
Segmentation precision is ensure that well, can effectively Ground Split background information be complicated and image of weak boundary.
Embodiments of the invention are the foregoing is only, are not intended to limit the scope of the invention, it is every to utilize this hair
The equivalents that bright specification and accompanying drawing content are made, or the technical field of correlation is directly or indirectly used in, similarly include
In the scope of patent protection of the present invention.
Claims (10)
1. a kind of image partition method, it is characterised in that including：
The salient region of target image is detected, the initialization boundary curve of target area is obtained；
According to the energy functional number of LIF models and the energy functional number of DRLSE models, new energy functional number is generated；
According to new the energy functional number and default iterations, the initialization boundary curve is developed, obtained
Boundary curve after evolution；
Image segmentation is carried out according to the boundary curve after the evolution.
2. image partition method according to claim 1, it is characterised in that " the conspicuousness area of detection target image
Domain, obtains the initialization boundary curve of target area " be specially：
According to cellular automata, the salient region of target image is detected, the initialization boundary curve of target area is obtained.
3. image partition method according to claim 2, it is characterised in that described " according to cellular automata, to detect target
The salient region of image, obtains the initialization boundary curve of target area " be specially：
Target image is divided into N number of super-pixel；
According to K mean algorithms, the super-pixel positioned at target image edge is divided into K classes；
According to the K class super-pixel after division, K global color disparity map and its corresponding notable value matrix are built；
The space length between super-pixel in each super-pixel and K class super-pixel, builds weight matrix；
According to the notable value matrix and weight matrix, calculating obtains notable figure；
According to the neighbouring relations and space length between super-pixel, factor of influence of the super-pixel to another super-pixel is defined, and
Build factor of influence matrix；
According to default degree matrix, the factor of influence matrix is normalized, the factor of influence matrix after being normalized；
According to the factor of influence in factor of influence matrix, the confidence level of each super-pixel current state is calculated, and builds confidence level square
Battle array；
The confidence level matrix is optimized, the confidence level matrix after being optimized；
According to the confidence level matrix after the factor of influence matrix after normalization and optimization, evolution renewal is carried out to the notable figure,
Obtain the final notable figure of target image；
According to the final notable figure, the initialization boundary curve of target area is obtained.
4. image partition method according to claim 1, it is characterised in that described " according to the energy functional number of LIF models
With the energy functional number of DRLSE models, new energy functional number is generated " be specially：
According to the energy functional number of LIF models and the energy functional number of DRLSE models, new energy functional number is generated, it is described new
Energy functional number beWherein, ELIFFor the energy functional number of LIF models, EDRLSEFor DRLSE
The energy functional number of model, η is the weight of default correspondence LIF models, and ρ is the weight of default correspondence DRLSE models；
According to the energy functional number of LIF models, the level set movements equation of LIF models is obtained；
According to the energy functional number of DRLSE models, the level set movements equation of DRLSE models is obtained；
According to the level set movements side of the new energy functional number, the level set movements equation of LIF models and DRLSE models
Journey, obtains new level set movements equation.
5. image partition method according to claim 4, it is characterised in that described " according to the new energy functional number
With default iterations, the initialization boundary curve is developed, the boundary curve after being developed " be specially：
According to new the level set movements equation and default iterations, the initialization boundary curve is developed,
Boundary curve after being developed.
6. a kind of image segmentation system, it is characterised in that including：
Detection module, the salient region for detecting target image obtains the initialization boundary curve of target area；
Generation module, for the energy functional number and the energy functional number of DRLSE models according to LIF models, generates new energy
Functional；
Genetic module, for according to new the energy functional number and default iterations, to the initialization boundary curve
Developed, the boundary curve after being developed；
Split module, for carrying out image segmentation according to the boundary curve after the evolution.
7. image segmentation system according to claim 6, it is characterised in that the detection module is specifically for according to cellular
Automatic machine, detects the salient region of target image, obtains the initialization boundary curve of target area.
8. image segmentation system according to claim 7, it is characterised in that the detection module includes：
First division unit, for target image to be divided into N number of super-pixel；
Second division unit, for according to K mean algorithms, the super-pixel positioned at target image edge to be divided into K classes；
First construction unit, for according to the K class super-pixel after division, building K global color disparity map and its corresponding showing
Write value matrix；
Second construction unit, for the space length between the super-pixel in each super-pixel and K class super-pixel, builds power
Weight matrix；
First computing unit, for according to the notable value matrix and weight matrix, calculating to obtain notable figure；
3rd construction unit, for according to the neighbouring relations and space length between super-pixel, defining a super-pixel to another super
The factor of influence of pixel, and build factor of influence matrix；
Normalization unit, for according to default degree matrix, being normalized, being obtained after normalization to the factor of influence matrix
Factor of influence matrix；
4th construction unit, for the factor of influence in factor of influence matrix, calculates the confidence of each super-pixel current state
Degree, and build confidence level matrix；
Optimize unit, for being optimized to the confidence level matrix, the confidence level matrix after being optimized；
Evolution updating block, for according to the confidence level matrix after the factor of influence matrix after normalization and optimization, to described aobvious
Write figure and carry out evolution renewal, obtain the final notable figure of target image；
First obtains unit, for according to the final notable figure, obtaining the initialization boundary curve of target area.
9. image segmentation system according to claim 6, it is characterised in that the generation module includes：
Generation unit, for the energy functional number and the energy functional number of DRLSE models according to LIF models, generates new energy
Functional, the new energy functional number isWherein, ELIFFor the energy functional number of LIF models,
EDRLSEFor the energy functional number of DRLSE models, η is the weight of default correspondence LIF models, and ρ is default correspondence DRLSE models
Weight；
Second obtains unit, for the energy functional number according to LIF models, obtains the level set movements equation of LIF models；
3rd obtains unit, for the energy functional number according to DRLSE models, obtains the level set movements equation of DRLSE models；
4th obtains unit, for the level set movements equation and DRLSE moulds according to the new energy functional number, LIF models
The level set movements equation of type, obtains new level set movements equation.
10. image segmentation system according to claim 9, it is characterised in that the genetic module is specifically for according to institute
New level set movements equation and default iterations is stated, the initialization boundary curve is developed, obtained after evolution
Boundary curve.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710404301.4A CN107330897B (en) | 2017-06-01 | 2017-06-01 | Image segmentation method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710404301.4A CN107330897B (en) | 2017-06-01 | 2017-06-01 | Image segmentation method and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107330897A true CN107330897A (en) | 2017-11-07 |
| CN107330897B CN107330897B (en) | 2020-09-04 |
Family
ID=60194046
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710404301.4A Active CN107330897B (en) | 2017-06-01 | 2017-06-01 | Image segmentation method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107330897B (en) |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108363055A (en) * | 2018-01-17 | 2018-08-03 | 电子科技大学 | A kind of radar foresight imaging region dividing method |
| CN108898611A (en) * | 2018-06-19 | 2018-11-27 | 东华理工大学 | Fuzzy region active contour parted pattern based on significant perception priori |
| CN110176021A (en) * | 2019-05-21 | 2019-08-27 | 山东大学 | In conjunction with the level set image segmentation method and system of the conspicuousness information of gamma correction |
| CN110288581A (en) * | 2019-06-26 | 2019-09-27 | 电子科技大学 | A kind of dividing method based on holding shape convexity Level Set Models |
| CN110390667A (en) * | 2019-06-18 | 2019-10-29 | 平安科技（深圳）有限公司 | Lesion extracting method, device, equipment and storage medium based on eyeground OCT image |
| WO2020029064A1 (en) * | 2018-08-07 | 2020-02-13 | 温州医科大学 | Optical coherence tomographic image processing method |
| CN111640188A (en) * | 2020-05-29 | 2020-09-08 | 中国地质大学（武汉） | Anti-noise three-dimensional grid optimization method based on Mumford-Shah algorithm framework |
| CN111739047A (en) * | 2020-06-24 | 2020-10-02 | 山东财经大学 | Tongue image segmentation method and system based on bispectrum reconstruction |
| CN112001933A (en) * | 2020-09-09 | 2020-11-27 | 成都市精卫鸟科技有限责任公司 | Image capturing method, device, equipment and medium |
| CN112037109A (en) * | 2020-07-15 | 2020-12-04 | 北京神鹰城讯科技股份有限公司 | Improved image watermarking method and system based on saliency target detection |
| CN112070712A (en) * | 2020-06-05 | 2020-12-11 | 宁波大学 | Printing defect detection method based on self-encoder network |
| CN112435263A (en) * | 2020-10-30 | 2021-03-02 | 苏州瑞派宁科技有限公司 | Medical image segmentation method, device, equipment, system and computer storage medium |
| CN113313672A (en) * | 2021-04-28 | 2021-08-27 | 贵州电网有限责任公司 | Active contour model image segmentation method based on SLIC superpixel segmentation and saliency detection algorithm |
| CN114445443A (en) * | 2022-01-24 | 2022-05-06 | 山东省人工智能研究院 | Interactive image segmentation method based on asymmetric geodesic distance |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102831608A (en) * | 2012-08-06 | 2012-12-19 | 哈尔滨工业大学 | Unsteady measurement algorithm based image segmentation method of improved rule distance level set |
| CN103093474A (en) * | 2013-01-28 | 2013-05-08 | 电子科技大学 | Three-dimensional mammary gland ultrasound image partition method based on homoplasmon and partial energy |
| CN103700095A (en) * | 2013-12-10 | 2014-04-02 | 东北林业大学 | Log end surface image partitioning algorithm for improving active contour model based on circle constraint |
| US20140285486A1 (en) * | 2013-03-20 | 2014-09-25 | Siemens Product Lifecycle Management Software Inc. | Image-based 3d panorama |
| CN104867143A (en) * | 2015-05-15 | 2015-08-26 | 东华理工大学 | Level set image segmentation method based on local guide core-fitting energy model |
| CN105825513A (en) * | 2016-03-21 | 2016-08-03 | 辽宁师范大学 | Image segmentation method based on adaptive fitting of global information and local information |
| CN106204592A (en) * | 2016-07-12 | 2016-12-07 | 东北大学 | A kind of image level collection dividing method based on local gray level cluster feature |
| CN106296649A (en) * | 2016-07-21 | 2017-01-04 | 北京理工大学 | A kind of texture image segmenting method based on Level Set Models |
| CN106548478A (en) * | 2016-10-28 | 2017-03-29 | 中国科学院苏州生物医学工程技术研究所 | Active contour image partition method based on local fit image |
-
2017
- 2017-06-01 CN CN201710404301.4A patent/CN107330897B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102831608A (en) * | 2012-08-06 | 2012-12-19 | 哈尔滨工业大学 | Unsteady measurement algorithm based image segmentation method of improved rule distance level set |
| CN103093474A (en) * | 2013-01-28 | 2013-05-08 | 电子科技大学 | Three-dimensional mammary gland ultrasound image partition method based on homoplasmon and partial energy |
| US20140285486A1 (en) * | 2013-03-20 | 2014-09-25 | Siemens Product Lifecycle Management Software Inc. | Image-based 3d panorama |
| CN103700095A (en) * | 2013-12-10 | 2014-04-02 | 东北林业大学 | Log end surface image partitioning algorithm for improving active contour model based on circle constraint |
| CN104867143A (en) * | 2015-05-15 | 2015-08-26 | 东华理工大学 | Level set image segmentation method based on local guide core-fitting energy model |
| CN105825513A (en) * | 2016-03-21 | 2016-08-03 | 辽宁师范大学 | Image segmentation method based on adaptive fitting of global information and local information |
| CN106204592A (en) * | 2016-07-12 | 2016-12-07 | 东北大学 | A kind of image level collection dividing method based on local gray level cluster feature |
| CN106296649A (en) * | 2016-07-21 | 2017-01-04 | 北京理工大学 | A kind of texture image segmenting method based on Level Set Models |
| CN106548478A (en) * | 2016-10-28 | 2017-03-29 | 中国科学院苏州生物医学工程技术研究所 | Active contour image partition method based on local fit image |
Non-Patent Citations (3)
| Title |
|---|
| CHUNMING LI ET AL;: "《Distance Regularized Level Set Evolution and Its Application to Image Segmentation》", 《IEEE TRANSACTIONS ON IMAGE PROCESSING》 * |
| 张迎春 等；: "《基于粗糙集和新能量公式的水平集图像分割》", 《自动化学报》 * |
| 袁建军: "《基于偏微分方程图像分割技术的研究》", 《博士学位论文库》 * |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108363055A (en) * | 2018-01-17 | 2018-08-03 | 电子科技大学 | A kind of radar foresight imaging region dividing method |
| CN108898611A (en) * | 2018-06-19 | 2018-11-27 | 东华理工大学 | Fuzzy region active contour parted pattern based on significant perception priori |
| WO2020029064A1 (en) * | 2018-08-07 | 2020-02-13 | 温州医科大学 | Optical coherence tomographic image processing method |
| CN110176021A (en) * | 2019-05-21 | 2019-08-27 | 山东大学 | In conjunction with the level set image segmentation method and system of the conspicuousness information of gamma correction |
| CN110176021B (en) * | 2019-05-21 | 2021-04-16 | 山东大学 | Level set image segmentation method and system for saliency information combined with brightness correction |
| CN110390667A (en) * | 2019-06-18 | 2019-10-29 | 平安科技（深圳）有限公司 | Lesion extracting method, device, equipment and storage medium based on eyeground OCT image |
| CN110390667B (en) * | 2019-06-18 | 2023-10-20 | 平安科技（深圳）有限公司 | Focus extraction method, device, equipment and storage medium based on fundus OCT image |
| CN110288581A (en) * | 2019-06-26 | 2019-09-27 | 电子科技大学 | A kind of dividing method based on holding shape convexity Level Set Models |
| CN110288581B (en) * | 2019-06-26 | 2022-11-04 | 电子科技大学 | Segmentation method based on model for keeping shape convexity level set |
| CN111640188A (en) * | 2020-05-29 | 2020-09-08 | 中国地质大学（武汉） | Anti-noise three-dimensional grid optimization method based on Mumford-Shah algorithm framework |
| CN112070712A (en) * | 2020-06-05 | 2020-12-11 | 宁波大学 | Printing defect detection method based on self-encoder network |
| CN112070712B (en) * | 2020-06-05 | 2024-05-03 | 湖北金三峡印务有限公司 | Printing defect detection method based on self-encoder network |
| CN111739047A (en) * | 2020-06-24 | 2020-10-02 | 山东财经大学 | Tongue image segmentation method and system based on bispectrum reconstruction |
| CN112037109A (en) * | 2020-07-15 | 2020-12-04 | 北京神鹰城讯科技股份有限公司 | Improved image watermarking method and system based on saliency target detection |
| CN112001933A (en) * | 2020-09-09 | 2020-11-27 | 成都市精卫鸟科技有限责任公司 | Image capturing method, device, equipment and medium |
| CN112435263A (en) * | 2020-10-30 | 2021-03-02 | 苏州瑞派宁科技有限公司 | Medical image segmentation method, device, equipment, system and computer storage medium |
| CN113313672A (en) * | 2021-04-28 | 2021-08-27 | 贵州电网有限责任公司 | Active contour model image segmentation method based on SLIC superpixel segmentation and saliency detection algorithm |
| CN114445443A (en) * | 2022-01-24 | 2022-05-06 | 山东省人工智能研究院 | Interactive image segmentation method based on asymmetric geodesic distance |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107330897B (en) | 2020-09-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107330897A (en) | Image partition method and its system | |
| CN105139015B (en) | A kind of remote sensing images Clean water withdraw method | |
| CN108319957A (en) | A kind of large-scale point cloud semantic segmentation method based on overtrick figure | |
| CN108764281A (en) | A kind of image classification method learning across task depth network based on semi-supervised step certainly | |
| CN108875595A (en) | A kind of Driving Scene object detection method merged based on deep learning and multilayer feature | |
| CN105513070A (en) | RGB-D salient object detection method based on foreground and background optimization | |
| CN108961675A (en) | Fall detection method based on convolutional neural networks | |
| CN109613006A (en) | A kind of fabric defect detection method based on end-to-end neural network | |
| CN104732545B (en) | The texture image segmenting method with quick spectral clustering is propagated with reference to sparse neighbour | |
| CN105740915B (en) | A kind of collaboration dividing method merging perception information | |
| CN110443809A (en) | Structure sensitive property color images super-pixel method with boundary constraint | |
| CN109101938A (en) | A kind of multi-tag age estimation method based on convolutional neural networks | |
| CN104021255A (en) | Multi-resolution hierarchical presenting and hierarchical matching weighted comparison method for CAD (computer aided design) model | |
| CN106485743A (en) | The functional assay of determinand threedimensional model and device | |
| CN110909778B (en) | Image semantic feature matching method based on geometric consistency | |
| CN105224935A (en) | A kind of real-time face key point localization method based on Android platform | |
| CN109829353A (en) | A kind of facial image stylizing method based on space constraint | |
| CN106022359A (en) | Fuzzy entropy space clustering analysis method based on orderly information entropy | |
| CN108846845B (en) | SAR image segmentation method based on thumbnail and hierarchical fuzzy clustering | |
| CN114140485A (en) | Method and system for generating cutting track of main root of panax notoginseng | |
| CN109872343A (en) | Weak texture gestures of object tracking, system and device | |
| JP7310912B2 (en) | 3D point cloud label learning device, 3D point cloud label estimation device, method, and program | |
| CN116778386A (en) | Sulfur hexafluoride leakage detection method, device, equipment and storage medium | |
| CN106251364A (en) | Method for tracking target and device | |
| CN102855624B (en) | A kind of image partition method based on broad sense data fields and Ncut algorithm |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20220411 Address after: 350000 B505, 3 / F, building 10, phase I, innovation park, No. 3, Keji East Road, high tech Zone, Fuzhou, Fujian Patentee after: Fujian Leji Technology Co.,Ltd. Address before: Science and Technology Office of Fujian Normal University Patentee before: Fujian Normal University |