CN101436255B - Method for extracting remarkable configuration in complicated image - Google Patents

Abstract

The invention discloses a method for extracting a remarkable outline from a complex image, which comprises the following steps: (1) an ideal outline is abstracted into a small line segment set, wherein small line segments are connected end to end, and the ideal outline is approached by the small line segment set; (2) a mathematical model for a line segment set W is established, wherein an optimum solution meeting the model is to approach the remarkable outline in the image; and (3) the optimum line segment set W is solved by applying reversible jump Markov chain Monte Carlo and simulated annealing algorithm. The method directly describes a unit (line segment) consisting of 'superior' outlines, avoids defect of 'noise' edges in the prior edge detection algorithm, gathers and unifies detection and perception of an outline composition unit under one framework, and finishes the detection and the perception simultaneously. The method avoids a perception gathering process form depending on the detection algorithm for the outline composition unit, and an output result is an approach to the ideal outline. An approach result has good anti-interference performance, and can obtain favorable vision effect for extracting the remarkable outline of the complex image.

Description

Method for extracting remarkable configuration in a kind of complicated image

Technical field

The invention belongs to computer vision field, be specifically related to method for extracting remarkable configuration in a kind of complicated image.

Background technology

The extraction of remarkable configuration is basic key problem in the computer vision in the image.For piece image, remarkable configuration is generally the geometric configuration of object in the image, or is used to distinguish the border between the object, and for example, Fig. 1 (c) is the result that remarkable configuration among Fig. 1 (a) is manually marked.People often just can identify by the perception to these remarkable configurations those objects in the image, and does not need too many information about object color and texture.Remarkable configuration is except being used for perception and identification, in computer vision, can also be used for compression, coupling, target following of image etc.

Usually, the process that remarkable configuration extracts is made up of two steps: the first step is the definition and the detection method of profile component units, algorithm commonly used is a component units of single pixel being regarded as profile, the method that detects, exactly by the pixel value in the single pixel local neighborhood is carried out filtering, and obtain the degree that this pixel belongs to profile, as, gradient calculation, methods such as Gabor filtering.In recent years, along with the widespread use of machine Learning Theory in computer vision, being used in profile based on the machine learning method that supervision is arranged detects, for example, the machine learning method that David R.Martin has adopted logic-based to return, scheme the machine learning method of tall and erect literary grace based on adaboost, the two has all provided the probable value that each pixel in the image belongs to profile.Above detection method is the contour extraction method based on pixel.Second step was from the result of first step output, extracted the profile with perception conspicuousness.In the piece image, if there is not noise, target object is artificial, and object shows does not have complicated texture, and the background at object place also is uniformly, so, we do not need the work in second step, only the output result with algorithms most in use in the first step just can obtain the significant profile of perception, for example, uses the most classical Canny edge detection algorithm.Yet, the image of the daily contact of people, most of under the natural scene, there is the noise of image, the irregular grain of body surface, complicated background etc.At present, in the computer vision, all have two problems as the profile testing method based on pixel in the above-mentioned first step, that is, to " fail to judge and judge by accident " of pixel, the pixel that has should be considered on the remarkable configuration, but is considered to not be; And the pixel that has should not be on the remarkable configuration, but is considered to, and for example, Fig. 1 (b) is exactly a kind of result who adopts the Canny edge detection algorithm to obtain, above-mentioned " fail to judge and judge by accident ", and situation has caused detected marginal existence " noise ".Certainly, the criterion between this " fail to judge and judge by accident " should be done final judgement by human eye.If the remarkable configuration that human eye perceived is thought final desired result, for example Fig. 1 (c) is exactly a kind of result that human eye is judged, the result who exports in the so above-mentioned first step and the gap of this desired result are exactly will make great efforts to reduce in above-mentioned second step.

At present, at the target that above-mentioned second step will finish, mainly be selectively to be gathered into remarkable configuration to the profile among the first step result.The perceptual elements that provides in the Gestalt psychology is provided the foundation of assembling usually.In present certain methods, usually directly pixel is not directly assembled, but assemble " senior " unit of some non-pixels, generally, this " senior " unit is little straight-line segment, so just can calculate direction between the line segment, distance etc. easily, can be more convenient utilize perceptual elements in the Gestalt psychology, as, proximity, collinearity etc.Usually some pre-service are done to the result of the first step by elder generation, such as, remove the pixel of " erroneous judgement ", these pixels can be chosen by the length of its place profile, usually in the result of the first step, length is less than the profile of certain threshold value, can be regarded as " erroneous judgement ", and these pixels are removed.On the other hand, fill the pixel of " failing to judge ", the simplest method is that each profile end points k neighborhood nearest with it end points is coupled together with straight-line segment.After finishing above-mentioned and handling, next be exactly that key element is assembled in the perception that utilizes Gestalt psychology to provide, define a kind of gathering standard, algorithm for design (at present algorithm adopt dynamic programming or global search) then more, searching out a subclass that satisfies the gathering standard from above-mentioned little line-segment sets comes, this subclass has constituted the final result of algorithm, it is the conspicuousness structure, it remains approaches a kind of of desired result, but compare with simple edge detection algorithm, it has reduced a lot " noise " edges, has kept most of structure of remarkable configuration, and using for follow-up computer vision provides input preferably.

Similar these above-mentioned methods, though obtained some comparatively satisfied results, but still exist not enough, the algorithm instability is a main defective, for simple relatively image, arithmetic result is satisfied, but after running into complex image, result often deviation is bigger, and its reason is, the result of rim detection in the searching algorithm dependence in above-mentioned second step and the above-mentioned first step, if the result of the first step is undesirable, after second step of input, even the algorithm in second step is best, also be difficult to obtain ideal results, as Stahl and Wang at document (J.S.Stahl, and S.Wang. " Edge Grouping Combining Boundary and Region Information ", IEEE Trans.ImageProcess., 16 (10): 377-384, mention in oct.2007.).

Summary of the invention

The object of the present invention is to provide method for extracting remarkable configuration in a kind of complicated image, this method is extracted for the remarkable configuration of complicated image, can obtain good visual effect.

Method for extracting remarkable configuration in a kind of complicated image provided by the invention the steps include:

(1) (x y), at first obtains to generate at random line segment center point coordinate p according to formula (i) for input picture I _i=(x, sampling probability q (p y) _i), the sequence number of pixel in the i presentation video:

$q\left(p_i\right)=\▿I(x,y)/{\mathrm{\Σ}}_{x,y}\▿I(x,y)---\left(i\right)$

Wherein

Be the Grad of image, its computing formula be formula (ii):

$\▿I(x,y)=\sqrt[2]{{[I(x+1,y)-I(x,y)]}^2+{[I(x,y+1)-I(x,y)]}^2}---\left(\mathrm{ii}\right)$

(2) set up line segment aggregate W mathematical model according to following process, to approaching of remarkable configuration in the image:

(2.1) initial value of set W is made as empty set, W={};

(2.2) the sampling probability q (p that obtains according to step (1) _i), sampling obtains line segment s ₀Center point coordinate p ₀=(x ₀, y ₀), then at default numerical value interval [l _Min, l _Max] on carry out uniform sampling, the length value l of this line segment of extraction ₀, next on direction interval [0, π], carry out uniform sampling, the direction value θ of this line segment of extraction ₀, with this line segment s ₀=(p ₀, l ₀, θ ₀) add and gather W;

(2.3) generate the random number rand between 0 to 1, record iterations N=N+1, the initial value of N is 0; If rand is smaller or equal to q _b, enter step (2.4), if rand is greater than q _bAnd smaller or equal to q _d+ q _b, change step (2.6) over to, if rand is greater than q _d+ q _b,, change step (2.8) over to; q _dBe the probability of the transfer form of deleting a line segment at random, q _bProbability for this transfer form that generates a line segment at random;

(2.4) use the method in the step (2.2) to generate a new line segment s at random _i=(p _i, l _i, θ _i), with its temporary transient set W that adds, at this moment, the new set of definition W is W ', W '=W ∪ { s _i, P () is a transition probability, it is defined as follows:

$P(w\&prime;,w)={\mathrm{<figure-callout\; id="1"\; label="q\; d"\; filenames="H2008102366036E00021.png,H2008102366036E00032.png"\; state="\{\{state\}\}">q</figure-callout>}}_d\frac{1}{\mathrm{card}\left(W\right)+1}---\left(\mathrm{iii}\right)$

P(w，w′)＝q _bq(p _i)q(l _i|p _i)q(θ _i|p _i) (iv)

The (iii) middle q of formula _dFor deleting the probability of a this transfer form of line segment at random, card (W) is the number of current state lower line segment; Q (p _i) for generating the sampling probability of line segment center point coordinate at random, it distributes and is determined by the gradient of image; Q (l _i| p _i) for generating the sampling probability of line segment length at random, it is distributed as even distribution; Q (θ _i| p _i) for generating the sampling probability of line segment direction at random, it is distributed as even distribution;

(2.5) (v) calculate line segment s according to formula _iProbability of acceptance α ₁(w, w '); Try to achieve line segment s _iProbability of acceptance α ₁(w, w ') afterwards, if the random number r between one 0 to 1 of the regeneration is α ₁(w, w ') greater than r, then line segment s _iTo finally be added to set W, otherwise will be lost; Change step (3) then over to;

(2.6) in current line segment aggregate W, line segment s of selected at random _i=(p _i, l _i, θ _i), it is temporarily deleted from set W; At this moment, W is W ' according to the new set of definition, deletes a line segment from current line segment aggregate W at random, this moment W '=W s _i, P () still adopts formula (iii) and (iv) to calculate;

(2.7) (v) calculate line segment s according to formula _iProbability of erasure α ₂(w, w ') tries to achieve line segment s _iProbability of erasure α ₂(w, w ') afterwards, if the random number r between one 0 to 1 of the regeneration is α ₂(w, w ') greater than r, then line segment s _iTo finally from set W, be deleted, otherwise will be not deleted; Change step (3) then over to;

(2.8) in current line segment aggregate W, line segment s of selected at random _i=(p _i, l _i, θ _i), if the random number a between generating 0 to 1 is a＜1/3, then l _i=l _i+ Δ l; If 1/3＜a＜2/3, then l _i=l _i-Δ l; If a〉2/3, l then _iRemain unchanged; Use the same method and adjust θ _iSize; Line segment center point coordinate p _iRemain unchanged; S like this _iTo become a new line segment s _i', the new set of same definition W is W ', W '=(W s _i) ∪ { s _i', s wherein _i'=(p _i, l _i± Δ l, θ _i± Δ θ), under this transfer form, P (w ', w) equal P (w, w ');

(2.9) (v) calculate line segment s according to formula _iAdjustment probability of acceptance α ₃(w, w '); Try to achieve line segment s _iAdjustment probability of acceptance α ₃(w, w ') afterwards, if the random number r between one 0 to 1 of the regeneration is α ₃(w, w ') greater than r, then line segment s _iS will finally be adjusted into _i', otherwise will remain unchanged;

α(w，w′)＝min{1，R(w，w′)} (v)

Wherein R (w, w ') is a Green ratio, and it is defined as:

$R(w,w\&prime;)=\frac{P(w\&prime;,w)f{(w\&prime;)}^{\frac{1}{T}}}{P(w,w\&prime;)f{\left(w\right)}^{\frac{1}{T}}}---\left(\mathrm{vi}\right)$

Wherein P () is a transition probability, and f () is the mathematical model of line segment aggregate, and T is the Current Temperatures in the simulated annealing;

(3) repeating step (2.3), every iteration once, the numerical value of N adds 1; When temperature T is tending towards 0, algorithm convergence, the line segment aggregate W of this moment is the optimum solution of this model, obtains the remarkable configuration of image.

The present invention is directed to the deficiency of contour extraction method in the existing complicated image, proposed a kind of mathematical model that remarkable configuration is approached, this model is assembled finishing the detection and the perception of showing the profile component units simultaneously.Directly to the profile component units of " senior ", line segment is described in the present invention, has avoided the defective at " noise " edge in the present edge detection algorithm, the detection of profile component units and perception is assembled being unified under the framework, finishes the realization of the two simultaneously.The inventive method has also been avoided the dependence of perception accumulation process to profile component units detection algorithm, and the result of its output approaches a kind of of desirable profile.The result who approaches has good anti-interference, and the remarkable configuration extraction for complicated image can obtain good visual effect.

Description of drawings

Fig. 1 shows natural image, the desirable profile of profile that obtains with the Canny rim detection and artificial mark;

Fig. 2 shows the algorithm realization flow figure of this method;

Fig. 3 shows a desirable profile and a line segment aggregate approaching it;

Fig. 4 (a) shows the attraction zone of middle conductor of the present invention; Fig. 4 (b) shows the connection situation of three line segments; Fig. 4 (c) shows the repulsive area of middle conductor of the present invention.

The width of cloth composograph that Fig. 5 (a) shows, the result of Fig. 5 (b) Canny edge detection algorithm, the result that Fig. 5 (c) the present invention obtains.

The width of cloth composograph that Fig. 6 (a) (b) shows, Fig. 6 (c) is the result of Canny edge detection algorithm (d), and Fig. 6 (e) is the result that obtains of the present invention (f).

Embodiment

The present invention is further detailed explanation below in conjunction with accompanying drawing and example.

As shown in Figure 2, the inventive method may further comprise the steps:

(1) (x y), at first obtains to generate at random line segment center point coordinate p for input picture I _i=(x, sampling probability q (p y) _i), the sequence number of pixel in the i presentation video:

$q\left(p_i\right)=\&dtri;I(x,y)/{\mathrm{\&Sigma;}}_{x,y}\&dtri;I(x,y).---\left(1\right)$

Wherein

For the Grad computing formula of image is:

$\&dtri;I(x,y)=\sqrt[2]{{[I(x+1,y)-I(x,y)]}^2+{[I(x,y+1)-I(x,y)]}^2}$

(2) set up line segment aggregate W mathematical model, satisfy the optimum solution of this model, be approaching remarkable configuration in the image.The present invention has used reversible jump Markov chain Monte-Carlo (Reversible Jump Markov chain Monte Carlo (RJMCMC)) and simulated annealing that optimum line segment set is found the solution.Its concrete steps are as follows:

(2.1) initial value of set W is made as empty set, W={};

(2.2) according to the probability distribution q (p in the formula (1) _i), carrying out stochastic sampling, a little line segment s samples out ₀Center point coordinate p ₀=(x ₀, y ₀).Then at numerical value interval [l _Min, l _Max] on carry out uniform sampling, the length value l of this little line segment of extraction ₀Next on direction interval [0, π], carry out uniform sampling, the direction value θ of this little line segment of extraction ₀, with this line segment s ₀=(p ₀, l ₀, θ ₀) add and gather W.

Numerical value interval [l _Min, l _Max] choose l according to experience _MinSpan be generally 7-9, l _MaxSpan be generally 15-17.If line segment length is too small, then calculated amount increases, and noise may occur simultaneously in the result; Line segment length is long, then can influence approaching of profile in the final image, approaches the not enough continuously smooth of result.

(2.3) generate the random number rand between 0 to 1.Record iterations N=N+1, the initial value of N is 0.

If (2.3.1) rand is less than 1/3, use the method in the step (2.2) to generate a new line segment s at random _i=(p _i, l _i, θ _i), with its temporary transient set W that adds.At this moment, the new set of definition W is W ', calculates line segment s then _iThe probability of acceptance, probability of acceptance computing formula is:

α(w，w′)＝min{1，R(w，w′)} (2)

Wherein R (w, w ') is a Green ratio, and it is defined as:

$R(w,w\&prime;)=\frac{P(w\&prime;,w)f{(w\&prime;)}^{\frac{1}{T}}}{P(w,w\&prime;)f{\left(w\right)}^{\frac{1}{T}}}---\left(3\right)$

Wherein P () is a transition probability, and f () is the Current Temperatures in the simulated annealing for the mathematical model of line segment aggregate (with describing in detail in the literal below), T.T reduces with the increase of iterations N, when N less than 2000 the time, T=T ₀, N is greater than after 2000, and the computing formula of T is:

$T=\frac{T_0}{\log (1+k\&times;2000)}---\left(4\right)$

K=floor (N/2000) wherein, that is, and less than the maximum integer of N divided by 2000 gained merchants.

Try to achieve line segment s _iProbability of acceptance α (w, w ') afterwards, the random number r between one 0 to 1 of the regeneration, if α (w, w ') greater than r, line segment s then _iTo finally be added to set W, otherwise will be lost.

If (2.3.2) rand is greater than 1/3 and less than 2/3, then in current line segment aggregate W, line segment s of selected at random _i=(p _i, l _i, θ _i), it is temporarily deleted from set W.At this moment, the new set of definition W is W ', calculates line segment s then _iProbability of erasure α (w, w '), computing formula still adopts formula (2).Try to achieve line segment s _iProbability of erasure α (w, w ') afterwards, the random number r between one 0 to 1 of the regeneration, if α (w, w ') greater than r, line segment s then _iTo finally from set W, be deleted, otherwise will be not deleted.

If (2.3.3) rand is greater than 2/3, then in current line segment aggregate W, line segment s of selected at random _i=(p _i, l _i, θ _i), if the random number a between generating 0 to 1 is a＜1/3, then l _i=l _i+ Δ l; If 1/3＜a＜2/3, then l _i=l _i-Δ l; If a〉2/3, l then _iRemain unchanged.Use the same method and adjust θ _iSize.Has only line segment center point coordinate p _iDo not adjust.S like this _iTo become a new line segment s _i', the new set of same definition W is W ', calculates line segment s then _iAdjustment probability of acceptance α (w, w '), computing formula still adopts formula (2).Try to achieve line segment s _iAdjustment probability of acceptance α (w, w ') afterwards, the random number r between one 0 to 1 of the regeneration, if α (w, w ') greater than r, line segment s then _iS will finally be adjusted into _i', otherwise will remain unchanged.

(2.4) repeating step (2.3), every iteration once, the numerical value of N adds 1.When temperature T is tending towards 0, algorithm convergence, the line segment aggregate W of this moment is the result that asks.

Specifying P () in the formula (3) and f () below:

1. in step (2.3.1), generate a new line segment at random, and add among the W, at this moment W '=W ∪ { s _i, P () is a transition probability, it is defined as follows:

$P(w\&prime;,w)={\mathrm{<figure-callout\; id="1"\; label="q\; d"\; filenames="H2008102366036E00021.png,H2008102366036E00032.png"\; state="\{\{state\}\}">q</figure-callout>}}_d\frac{1}{\mathrm{card}\left(W\right)+1}---\left(5\right)$

P(w，w′)＝q _bq(p _i)q(l _i|p _i)q(θ _i|p _i) (6)

Formula (5), (6), card (W) is the number of current state lower line segment, q _dFor deleting the probability of a this transfer form of line segment at random, q in the formula _bBe the probability of this transfer form that generates a line segment at random, q (p _i) for generating the sampling probability of line segment center point coordinate at random, it distributes and is determined by the gradient of image; Q (l _i| p _i) for generating the sampling probability of line segment length at random, it is distributed as even distribution; Q (θ _i| p _i) for generating the sampling probability of line segment direction at random, it is distributed as even distribution.By the random number rand of step (2.3) beginning as can be known, q _bAnd q _dBe 1/3.

In step (2.3.2), from current line segment aggregate W, delete a line segment at random, this moment W '=W s _i, be an inverse process with step (2.3.1) in fact, P () still adopts formula (5) and (6) to calculate, and just w wherein is different with w '.

In step (2.3.3), line segment of picked at random in current line segment set W, then, its length of adjustment at random and direction again, this moment W '=(W s _i) ∪ { s _i', s wherein _i'=(p _i, l _i± Δ l, θ _i± Δ θ), under this transfer form, P (w ', w) equal P (w, w ').

At the mathematical model f () that line segment aggregate is described before, three types of line segment at first are described, shown in figure (3), each little line segment all has two end points, be assumed to be a U and put V, if the distance between the end points of end points of a little line segment and the little line segment of another one less than d, thinks then that these two little line segments are for linking to each other.In view of the above, little line segment can be divided into type in 3, and is continuous if two end points of a little line segment are, and claims that then this line segment is the doubly-linked line segment; If have only an end points to link to each other, then be called the hookup wire section; If two end points all do not link to each other, then be called free line segment.

(i) mathematical model of little line segment aggregate is:

f(W)∝β ⁿexp(-E(W))＝β ⁿexp-(E _I(W)+E _D(W)) (7)

This model has adopted the expression of probability form, and f (W) does not have normalized probability density formula, and β represents the density of ripple pine process, promptly is used for controlling the number of little line segment, and n represents the number of little line segment under the current state, E _I(W) associated energies between the expression line segment, it has described the relation between the little line segment space distribution under the current state.E _D(W) the data energy of expression line segment, it has described the fitting degree of little line segment to desirable profile in the real image.By above-mentioned modeling, provide piece image, to approaching of its remarkable configuration, obtain exactly and make in the formula (7) that f (W) is a maximum set W.Its mathematical notation is exactly:

W ^*＝arg?min{E _I(W)+E _D(W)-nlogβ} (8)

Among the present invention, will following formula-nlog β merges to E _I(W) in.

(ii) be used for describing the E of line segment space distribution geometric relationship _I(W) be defined as:

$E_1\left(W\right)={\mathrm{\&omega;}}_{0}n+{\mathrm{\&omega;}}_1{n}_f+{\mathrm{\&omega;}}_2{n}_s+{\mathrm{\&omega;}}_a{\mathrm{\&Sigma;}}_{<s_i,s_j>}{I}_a(s_i,s_j)+{\mathrm{\&omega;}}_r{\mathrm{\&Sigma;}}_{<s_i,s_j>}{I}_r(s_i,s_j)---\left(9\right)$

ω in the formula ₀Be exactly in the formula (3)-log β ω ₁, ω ₂, ω _r, ω _aRepresent every energy punishment parameter, except ω _aOutside negative, all the other are positive number.N represents current set middle conductor sum, n _fRepresent free line segment number in the current set, n _sRepresent hookup wire section number in the current set.＜s _i, s _jRepresent that two continuous line segments are right in the current set, I _a(s _i, s _j) attraction energy between the expression two-phase line section, I _r(s _i, s _j) repulsion energy between the expression two-phase line section.

The attraction that (iii) links to each other between the line segment is calculated as with the repulsion energy:

$I_a(s_i,s_j)=\underset{{\mathrm{<figure-callout\; id="1"\; label="O"\; filenames="H2008102366036E00021.png,H2008102366036E00032.png"\; state="\{\{state\}\}">O</figure-callout>}}_1=U_{s_i},V_{s_i}}{\mathrm{\&Sigma;}}\underset{O_2=U_{s_j},V_{s_j}}{\mathrm{\&Sigma;}}\frac{A(C_{O_1}^{l_{s_i}/4}\&cap;C_{O_2}^{l_{s_j}/4})}{\min (A\left(C_{O_1}^{l_{s_i}/4}\right),A\left(C_{O_2}^{l_{s_j}/4}\right))}$ (10)

$I_r(s_i,s_j)=\frac{A(C_{P_{s_i}}^{l_{s_i}/2}\&cap;C_{P_{s_j}}^{l_{s_j}2})}{\min (A\left(C_{P_{s_i}}^{l_{s_i}/2}\right),A\left(C_{P_{s_j}}^{l_{s_j}/2}\right))}$ (11)

To count the definition that two formulas have all adopted a border circular areas in

Be that the center of circle is the circle of r for the o radius.A (.) expression is to the calculating of area.The distribution of profile is at random in the image, the distribution that has caused little line segment also is at random, in this case, we wish that adjacent line segment joins end to end, and by the attraction zone of line segment, just can quantize the degree that links to each other and close between the line segment, as the dashed circle zone among Fig. 4 (a), its radius is 1/4 of a line segment length, and the center of circle is the end points of line segment, and clearly definition is arranged in formula (10).Yet, should satisfy the conllinear continuity between two line segments, that is, the too big difference that do not have of the direction between the line segment.As Fig. 4 (b), the linking to each other of three line segments, and share an end points, by the calculating of formula (10), line segment is right＜s ₁, s ₂With＜s ₁, s ₃Have identical attraction energy, and learn s from experience ₁With s ₃Between connection should be more reasonable, more identical with Gestalt psychology.Based on such consideration, formula (11) has been considered the repulsion factor between two line segments, and as Fig. 4 (c), broken circle is the repulsive area of three line segments, and its radius is 1/2 of a corresponding line segment, and the center of circle is that the mid point of line segment (is the P in the formula (11) _s).

(iv) be used to describe the E of line-segment sets data energy _D(W) be defined as:

$E_D\left(W\right)={\mathrm{\&omega;}}_d{\mathrm{\&Sigma;}}_{s_i\&Element;W}\mathrm{\&mu;}\&prime;\left(s_i\right)---\left(12\right)$

ω in the formula _dFor energy punishment parameter, get positive number.μ ' (s _i) be to the fitting degree μ (s of single line segment to profile _i) re-quantization, μ (s _i) definition, depend on the type of concrete image, such as, colour, two-value, gray level image etc. specifically are defined in the following instantiation and provide, in any case but definition μ (s _i), the present invention is to μ ' (s _i) definition be:

$\mathrm{\&mu;}\&prime;\left(s_i\right)=\left\{\begin{array}{cc}\&infin;& \mathrm{if\&mu;}\left(s_i\right)<{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="H2008102366036E00021.png,H2008102366036E00032.png"\; state="\{\{state\}\}">t</figure-callout>}}_1\\ 1-2\frac{\mathrm{\&mu;}\left(s_i\right)-t_1}{t_2-{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="H2008102366036E00021.png,H2008102366036E00032.png"\; state="\{\{state\}\}">t</figure-callout>}}_1}& \mathrm{if}{\mathrm{<figure-callout\; id="1"\; label="t"\; filenames="H2008102366036E00021.png,H2008102366036E00032.png"\; state="\{\{state\}\}">t</figure-callout>}}_1\&le;\mathrm{\&mu;}\left(s_i\right)<t_2\\ -1& \mathrm{if\&mu;}\left(s_i\right)\&GreaterEqual;t_2\end{array}\right.---\left(13\right)$

The profile of having realized synthetic bianry image and gray level image among the present invention extracts, and for synthetic bianry image, shown in Fig. 5 (a), its pixel value is 0 or 255, for this class image μ (s _i) numerical value be the pixel value average of all respective pixel on the line segment, in such cases, t among the present invention ₁=190, t ₂=225.

For real gray level image, as Fig. 6 (a) (b),, calculating μ (s for this class image _i) before, to each line segment, define two rectangles earlier, the size of these two rectangles is l * Δ w, l is the current length of line segment, Δ w=4, these two rectangles are their a shared limit with line segment, and they and line segment are in the same way, the two is respectively along the direction of line segment, be in the left side and the right of line segment, in such cases $\mathrm{\&mu;}\left(s_i\right)=|m_R-m_L|/\sqrt{{\mathrm{\&sigma;}}_R^2/n_R+{\mathrm{\&sigma;}}_L^2/n_L},$ m _R, m _LBe the average of pixel value in the right and the left side rectangle,

Be the standard deviation of correspondence, n _R, n _LBe the number of pixel in the right and the left side rectangle, the two equates here, in such cases, and t among the present invention ₁=9, t ₂=10.

Example:

(1) systematic parameter in the initialization formula (9) (12):

ω ₀＝20，ω ₁＝220，ω ₂＝40，ω _a＝-10000，ω _r＝10000，ω _d＝10

T ₀＝25

In addition, the width of line segment is 1 pixel, and length range is l _Min=9, l _Max=17

Length that line segment is adjusted at random and direction are Δ l=2, Δ θ=π/12

(2) (x y), at first obtains to generate at random line segment center point coordinate p for input picture I _i=(x, sampling probability y):

$q\left(p_i\right)=\&dtri;I(x,y)/{\mathrm{\&Sigma;}}_{x,y}\&dtri;I(x,y).$

Wherein

For the Grad computing formula of image is:

$\&dtri;I(x,y)=\sqrt[2]{{[I(x+1,y)-I(x,y)]}^2+{[I(x,y+1)-I(x,y)]}^2}.$

(3) adopt reversible jump Markov chain Monte-Carlo (Reversible Jump Markov chain Monte Carlo (RJMCMC)) algorithm to carry out iterative, iterations is 3*10 ⁵Inferior.

The result that Fig. 5 and Fig. 6 obtain for algorithm of the present invention.

According to an exemplary embodiment of the present invention, be used to realize that computer system of the present invention can comprise, particularly, central processing unit (CPU), storer and I/O (I/O) interface.Computer system usually by I/O interface and display with link to each other such as this type of various input equipments of mouse and keyboard, support circuit can comprise the fast buffer memory of image height, power supply, clock circuit and the such circuit of communication bus.Storer can comprise random access memory (RAM), ROM (read-only memory) (ROM), disc driver, magnetic tape station etc., or their combination.Computer platform also comprises operating system and micro-instruction code.Various process described herein and function can be by the micro-instruction code of operating system execution or the part of application program (or their combination).In addition, various other peripherals can be connected to this computer platform, as additional data storage device and printing device.

Should also be understood that and so the actual connection between the system component (or process steps) may be different, specifically decide on programming mode of the present invention because the assembly and the method step of some construction system described in the accompanying drawing can form of software be realized.Based on the principle of the invention that proposes herein, the ordinary skill of association area it is contemplated that these and similar embodiment or configuration of the present invention.

Claims (1)

1. method for extracting remarkable configuration in the complicated image, its step comprises:

(1) (x y), at first obtains to generate at random line segment center point coordinate p according to formula (i) for input picture I _i=(x, sampling probability q (p y) _i), the sequence number of pixel in the i presentation video:

q(p _i)＝▽I(x，y)/∑ _x，y▽I(x，y) (i)

Wherein ▽ I (x y) is the Grad of image, its computing formula be formula (ii):

$\&dtri;I(x,y)=\sqrt[2]{{[I(x+1,y)-I(x,y)]}^2+{[I(x,y+1)-I(x,y)]}^2}---\left(\mathrm{ii}\right)$

(2) set up line segment aggregate W mathematical model according to following all processes, remarkable configuration in the image approached:

(2.1) initial value of set W is made as empty set, W={};

(2.2) the sampling probability q (p that obtains according to step (1) _i), sampling obtains a line segment s ₀Center point coordinate p ₀=(x ₀, y ₀), then at default numerical value interval [l _Min, l _Max] on carry out uniform sampling, the length value l of this line segment of extraction ₀, next on direction interval [0, π], carry out uniform sampling, the direction value θ of this line segment of extraction ₀, with this line segment s ₀=(p ₀, l ₀, θ ₀) add and gather W;

(2.3) generate the random number rand between 0 to 1, record iterations N=N+1, the initial value of N is 0; If rand is smaller or equal to q _b, enter step (2.4), if rand is greater than q _bAnd smaller or equal to q _d+ q _b, change step (2.6) over to, if rand is greater than q _d+ q _b,, change step (2.8) over to; q _dBe the probability of the transfer form of deleting a line segment at random, q _bProbability for the transfer form that generates a line segment at random;

(2.4) use the method in the step (2.2) to generate a new line segment s at random _i=(p _i, l _i, θ _i), with its temporary transient set W that adds, at this moment, the new set of definition W is W ', W '=W ∪ { s _i, P () is a transition probability, it is defined as follows:

$P(w^{\&prime;},w)=q_d\frac{1}{\mathrm{card}\left(W\right)+1}---\left(\mathrm{iii}\right)$

P(w，w′)＝q _bq(p _i)q(l _i|p _i)q(θ _i|p _i) (iv)

The (iii) middle card (W) of formula is the number of current state lower line segment; Q (p _i) for generating the sampling probability of line segment center point coordinate at random, it distributes and is determined by the gradient of image; Q (l _i| p _i) be the sampling probability that generates line segment length at random, and q (l _i| p _i) it is distributed as even distribution; Q (θ _i| p _i) be the sampling probability that generates line segment direction at random, and q (θ _i| p _i) it is distributed as even distribution;

(2.5) (v) calculate line segment s according to formula _iProbability of acceptance α ₁(w, w '); Try to achieve line segment s _iProbability of acceptance α ₁(w, w ') afterwards, if the random number r between one 0 to 1 of the regeneration is α ₁(w, w ') greater than r, then line segment s _iTo finally be added to set W, otherwise will be lost; Change step (3) then over to;

(2.6) in current line segment aggregate W, line segment s of selected at random _i=(p _i, l _i, θ _i), it is temporarily deleted from set W; At this moment, W is W ' according to the new set of definition, deletes a line segment from current line segment aggregate W at random, this moment W '=W s _i, P () still adopts formula (iii) and (iv) to calculate;

(2.7) (v) calculate line segment s according to formula _iProbability of erasure α ₂(w, w ') tries to achieve line segment s _iProbability of erasure α ₂(w, w ') afterwards, if the random number r between one 0 to 1 of the regeneration is α ₂(w, w ') greater than r, then line segment s _iTo finally from set W, be deleted, otherwise will be not deleted; Change step (3) then over to;

(2.8) in current line segment aggregate W, line segment s of selected at random _i=(p _i, l _i, θ _i), if the random number a between generating 0 to l is a＜1/3, then l _i=l _i+ Δ l; If 1/3＜a＜2/3, then l _i=l _i-Δ l; If a＞2/3, then l _iRemain unchanged; Use the same method and adjust θ _iSize; Line segment center point coordinate p _iRemain unchanged; S like this _iTo become a new line segment s _i', the new set of same definition W is W ', W '=(W s _i) ∪ { s _i', s wherein _i'=(p _i, l _i± Δ l, θ _i± Δ θ), under this transfer form, P (w ', w) equal P (w, w ');

(2.9) (v) calculate line segment s according to formula _iAdjustment probability of acceptance α ₃(w, w '); Try to achieve line segment s _iAdjustment probability of acceptance α ₃(w, w ') afterwards, if the random number r between one 0 to 1 of the regeneration is α ₃(w, w ') greater than r, then line segment s _iS will finally be adjusted into _i', otherwise will remain unchanged;

α(w，w′)＝min{1，R(w，w′)} (v)

Wherein R (w, w ') is a Green ratio, and it is defined as:

$R(w,w^{\&prime;})=\frac{P(w^{\&prime;},w)f{\left(w^{\&prime;}\right)}^{\frac{1}{T}}}{P(w,w^{\&prime;})f{\left(w\right)}^{\frac{1}{T}}}---\left(\mathrm{vi}\right)$

Wherein f () is the mathematical model of line segment aggregate, and T is the Current Temperatures in the simulated annealing;

(3) repeating step (2.3), every iteration once, the numerical value of N adds 1; When temperature T is tending towards 0, algorithm convergence, the line segment aggregate W of this moment is the optimum solution of this model, obtains the remarkable configuration of image.

Images