CN103996185A - Image segmentation method based on attention TD-BU mechanism - Google Patents

Abstract

The invention discloses an image segmentation method based on an attention TD-BU mechanism. The method comprises the following steps: S1, a target image is acquired; S2, a nine-scale map is produced by the use of a binary Gauss Pyramid according to an input scene image and the target image; S3, scene feature maps are generated according to the scene image and the target image; S4, scene saliency maps are generated according to scene features, and the scene saliency maps are fused across scale to generate a final attention saliency map S; S5, the final attention saliency map S is interpolated in a bilinear mode to 0 scale and then binarized; and S6, bit-by-bit AND operation is performed on the binarized final attention saliency map S and an original input scene map, and a target object is segmented. The method put forward by the invention is a multi-scale multi-feature analysis method, namely, a data-driven bottom-up attention mechanism determined by object space association relationship saliency, and a task-driven top-down attention mechanism determined by target features. Brightness, color and direction image features and target color feature of a vision scene are jointly extracted in a multi-scale space to generate a scene salient feature map.

Description

A kind of image partition method based on notice TD-BU mechanism

Technical field

The invention belongs to image processing field, be specifically related to a kind of image partition method based on notice of combining target color character.

Background technology

Visual attention mechanism mainly comprises two types: notice (bottom-up attention from bottom to top, be abbreviated as " BU notice ") mechanism and notice (top-down attention is abbreviated as " TD notice ") mechanism from top to bottom.BU notice mechanism i.e. the process to upper strata notice by bottom data, notices that obtaining of target derives from scene image completely.TD notice mechanism is followed top-down process, decides the region of the power that attracts attention in lower floor's scene image according to task object.

Simple TD notice mechanism has independence and the complicacy of height, is mainly reflected in the uncertain and random of task object, therefore be difficult to set up more accurately its model mechanism, causes the limitation on using.On the other hand, BU notice process is carried out Selective attention power marking area according to the feature such as such as brightness, color and direction of input scene image, this process can be regarded as the attention selection mechanism of mankind's early vision system, therefore its modeling is relatively easy, receive the concern of Many researchers, representative is wherein Itti model.Its utilizes extracts the low-level features such as brightness, color and direction structure visual attention and significantly schemes from image, by the remarkable figure of the vision power that attracts attention, segmentation object object.The method model is simple, good to the robustness of noise, particularly in the time that the characteristic of noise is not directly destroyed the principal character of target.Shortcoming is that chaff interference also has significance in various degree in the time that chaff interference and target have Partial Feature similar, thereby the attention of target is produced to certain influence.Amudha.J. proposed a kind of new cognitive method based on notice with Soman.K.P., the method is noted region detection with the selecting tuning model of tasks in parallel and merging.When this model comprises preprocessing process as target identification method, effectively reduce the computation complexity of search, the test of using in the enterprising row labels of video sequence also obtains comparatively ideal result.L.Itti and P.Baldi propose the surprised model of a kind of new Bayes, and model applies to the surprised model of human visual attention power bayesian theory, weighs the attraction degree of data to observer with this.Although this surprised model is the low-level quantification on scene time and space just, there is no abundant semantic rules, mankind's area-of-interest also detected more accurately, just, compared with additive method, the interesting target detecting is relatively many.The people such as Wen.G have proposed a kind of visual attention sensor model based on wisp, this model, taking Itti attention model as basis, uses gauss hybrid models (GMM) to carry out further precision to remarkable figure and obtains area-of-interest after generating remarkable figure.The method can more accurately detect target, especially the perception of little target is had to good effect.

Several sensor models based on notice can be obtained preferably the sensing results of expection in specific utilization scene above, but through scrutinizing discovery, when the significance degree of target in scene is during lower than background or other objects, notice is not often attracted by target, thereby can not obtain satisfied result.

Summary of the invention

Given this, the present invention proposes a kind of improved image partition method based on notice, makes also to be cut apart more accurately during lower than background or other objects when target significance.

The object of the invention is to realize by such technical scheme, based on the image partition method of notice, it is characterized in that: comprise the following steps: S1 gathers target image; S2 uses two to enter gaussian pyramid and produce 9 scalograms the scene image of input and target image; S3 is according to scene image and target image generating scene characteristic pattern; S4 significantly schemes according to scene characteristic generating scene, remarkable scene figure is merged to the final notice of generation across yardstick and significantly scheme S; S5 significantly schemes S bilinear interpolation to 0 yardstick, then binaryzation to final notice; S6 by the final notice of binaryzation significantly scheme S and former input scene graph step-by-step phase with, be partitioned into target object.

Further, described scene characteristic figure comprises color characteristic figure, brightness figure, the direction character figure being extracted by data-driven and the object color component characteristic pattern extracting according to goal task.

Further, through type (2) calculates brightness figure I (c, s),

I(c,s)＝|I(c)ΘI(s)| (2)

Wherein, yardstick centered by c, s is yardstick around;

By r, g, b triple channel expands to R, G, B, tetra-passages of Y, are defined as follows,

$\left\{\begin{array}{c}R=r-(g+b)/2\\ G=g-(r+b)/2\\ B=b-(g+b)/2\\ Y=(r+b)-|r-b|/2-b\end{array}\right.---\left(3\right)$

R, g, b represents respectively three passages of red, green, blue in rgb color model, its brightness is I=(r+g+b)/3, gaussian pyramid R (δ), G (δ), B (δ), Y (δ) is obtained by above four new tunnels respectively;

Through type (4) calculates color characteristic figure,

$\left\{\begin{array}{c}\mathrm{RG}(c,s)=\left|(R\left(c\right)-G\left(c\right))\mathrm{\Θ}(G\left(s\right)-R\left(s\right))\right|\\ \mathrm{BY}(c,s)=\left|(B\left(c\right)-Y\left(c\right))\mathrm{\Θ}(Y\left(s\right)-B\left(s\right))\right|\end{array}\right.---\left(4\right)$

Wherein, RG (c, s) represents red and green and green and red two colors pair, and BY (c, s) represents blue and yellow and yellow and blue two colors pair;

Through type (5) calculated direction characteristic pattern O (c, s, θ),

O(c,s,θ)＝|O(c,θ)ΘO(s,θ)| (5)

Wherein, θ ∈ { 0 °, 45 °, 90 °, 135 ° } represents four direction,

Calculate by the following method object color component characteristic pattern P (c, s, m),

Computing Principle represents the point in scene image as divided into f (i, j), and f (i, j) corresponding point color value is (r ₀, g ₀, b ₀); The color value that m is corresponding is (r, g, b), and m ∈ (m1, m2): m1, m2 represent respectively two kinds of maximum colors of colour component proportion in target image, and scene image color is more similar to m, and corresponding point eigenwert is larger;

The too low corresponding point of similarity eigenwert is 0, even | and r ₀-r|or|g ₀– g|or|b ₀– b| is greater than T, and T is threshold value, P (ψ, m)=0; Otherwise

P(ψ，m)＝(1-|f(i,j)-m|) ³,ψ∈{c,s}

P(c,s,m)＝|P(c,m)ΘP(s,m)| (6)。

Further, the method that generating scene is significantly schemed S is first the scene characteristic figure generating in S2 to be fused to four significantly figure with normalization with across yardstick; The remarkable figure of this four width is integrated into final significantly figure S, the described remarkable figure of four width is respectively brightness significantly to scheme again color is significantly schemed direction is significantly schemed significantly scheme with object color component normalized computing method: first image value specification is turned to [0, M]; Secondly the mean value of maximal value M and other all local maximums in computed image and then entire image is multiplied by merging across yardstick is that each normalized characteristic pattern is compressed to yardstick 4 corresponding point addition again; The computing formula that final notice is significantly schemed S is as follows:

$\begin{array}{c}\tilde{I}=\underset{c=2}{\overset{4}{\⊕}}\underset{s=c+3}{\overset{c+4}{\⊕}}N\left(I(c,\tilde{s})\right)\\ \tilde{C}=\underset{c=2}{\overset{4}{\⊕}}\underset{s=c+3}{\overset{c+4}{\⊕}}[N\left(\mathrm{RG}(c,s)\right)+N\left(\mathrm{BY}(c,s)\right)]\\ \tilde{O}=\underset{\mathrm{\θ}=\left\{\mathrm{0,45,90,135}\right\}}{\mathrm{\Σ}}N\left(\underset{c=2}{\overset{4}{\⊕}}\underset{s=c+3}{\overset{c+4}{\⊕}}N\left(O(c,s,\mathrm{\θ})\right)\right)\\ \widetilde{P=}\underset{m=\{m1,m2\}}{\mathrm{\Σ}}N\left(\underset{c=2}{\overset{4}{\⊕}}\underset{s=c+3}{\overset{c+4}{\⊕}}N\left(P(c,s,m)\right)\right)\\ S=(\tilde{I}+\tilde{C}+\tilde{O}+\tilde{P})/4\end{array}---\left(7\right)$

Owing to having adopted technique scheme, the present invention has advantages of as follows:

A kind of new image partition method based on notice is proposed herein.The method is a kind of multiple dimensioned many characteristic analysis methods, has not only utilized scene information from bottom to top but also has utilized the color characteristic of goal task jointly to form scene image and significantly schemed, thereby then from remarkable figure, having extracted marking area and reach the object of perception.

Because the method has added goal task feature, make also to be cut apart more accurately during lower than background or other objects when target significance.

Brief description of the drawings

In order to make the object, technical solutions and advantages of the present invention clearer, below in conjunction with accompanying drawing, the present invention is described in further detail, wherein:

Fig. 1 is the process flow diagram of image partition method in the present invention;

Fig. 2 is the operation result figure improving one's methods;

Fig. 3 is visual sense feeling result comparison diagram;

Fig. 4 is to noise experiment result figure.

Embodiment

Below with reference to accompanying drawing, the preferred embodiments of the present invention are described in detail; Should be appreciated that preferred embodiment is only for the present invention is described, instead of in order to limit the scope of the invention.

Image partition method in this paper, mainly comprises the following steps:

S1 gathers target image.

S2 uses two to enter gaussian pyramid and produce 9 scalograms (0 scene image of input and target image ... 8 yardsticks), in 8 octaves from 1:1,1:2,1:4 ..., 1:256 (correspondence 0,1,2 respectively ..., 8 yardsticks).Calculate each feature by accepting wild similarly linear " center around " operation with vision, then use across yardstick and merge.Center around operator in model by around yardstick (s) interpolation Hou Yu center yardstick (c) carry out point-to-point subtracting each other accomplished (below representing with Θ).Center is that { pixel in 2,3,4}, is around yardstick s=c+d to yardstick c ∈, wherein d ∈ { 3,4}.

S3 is according to scene image and target image generating scene characteristic pattern.Wherein scene characteristic map generalization is made up of two category features: merely from bottom to top, and the object color component feature that three category features (color characteristic, brightness, direction character) that extracted by data-driven and combining target task are extracted.

S4 significantly schemes according to scene characteristic generating scene, remarkable scene figure is merged to the final notice of generation across yardstick and significantly scheme S.Four features that generate remarkable figure are respectively brightness, color characteristic, direction character and object color component feature.

For rgb color model (b represents respectively three passages of red, green, blue for r, g), brightness I by

I＝(r+g+b)/3 (1)

Obtain, produce 9 yardstick gaussian pyramid I (δ) by center around operation, δ ∈ [0..8] represents yardstick.R, g, the I normalization of b passage, object is to reduce the coupling of brightness to color.Due to only in the time that brightness acquires a certain degree, color could be perceived, thus in the time that being greater than whole brightness of image 1/10, I is normalized operation, otherwise r=g=b=0.

Brightness figure I (c, s) computing formula is as follows:

I(c,s)＝|I(c)ΘI(s)| (2)

By r, g, b triple channel expands to R, G, B, tetra-passages of Y, are defined as follows:

$\left\{\begin{array}{c}R=r-(g+b)/2\\ G=g-(r+b)/2\\ B=b-(g+b)/2\\ Y=(r+b)-|r-b|/2-b\end{array}\right.---\left(3\right)$

Gaussian pyramid R (δ), G (δ), B (δ), Y (δ) is obtained by above four new tunnels respectively.Characteristic pattern by gaussian pyramid carry out center around operation be mentioned above by careful yardstick (c) obtain with the difference of thick yardstick (s) around.

The structure of color characteristic figure is according to primate brain cortex " color two competition " system: at vision acceptance domain center, neuron suppresses another kind of color (such as green), vice-versa when being excited by a kind of color (such as redness).In mankind's primary visual cortex, just there are like this four spatial color pair: red and green, green and red, Lan Yuhuang, Huang Yulan.Use characteristic pattern RG (c, s) to represent red and green and green and red two colors pair herein, blue and yellow and Huang and blue two colors pair with characteristic pattern BY (c, s) expression, its computing formula is as follows:

$\left\{\begin{array}{c}\mathrm{RG}(c,s)=\left|(R\left(c\right)-G\left(c\right))\mathrm{\Θ}(G\left(s\right)-R\left(s\right))\right|\\ \mathrm{BY}(c,s)=\left|(B\left(c\right)-Y\left(c\right))\mathrm{\Θ}(Y\left(s\right)-B\left(s\right))\right|\end{array}\right.---\left(4\right)$

The structure of direction character figure has used direction Garbor pyramid O (δ, θ), and δ represents yardstick, and θ ∈ { 0 °, 45 °, 90 °, 135 ° } represents four direction.(Garbor wave filter uses cosine grating and dimensional Gaussian convolution, and object is consistent with neuron susceptibility in the visual field of main responsible directional perception in visual cortex).Direction character figure O (c, s, θ) computing formula is as follows:

O(c,s,θ)＝|O(c,θ)ΘO(s,θ)| (5)

The formation of object color component characteristic pattern is that (two kinds of color values are used respectively m1 according to two kinds of maximum colors of colour component proportion in input target image, m2 represents), object is the color characteristic of strengthening target, and making also can be perceived under its situation of disturbing at more significant background or other objects.Object color component characteristic pattern P (c, s, m), Computing Principle is following (establishes f (i, j) and represent the point in scene image, and f (i, j) corresponding point color value to be (r ₀, g ₀, b ₀); The color value that m is corresponding is (r, g, b), m ∈ (m1, m2)):

Scene image color is more similar to m, and corresponding point eigenwert is larger;

The too low corresponding point of similarity eigenwert is 0, even | and r ₀-r|or|g ₀– g|or|b ₀– b| is greater than T, P (ψ, m)=0; Otherwise

P(ψ，m)＝(1-|f(i,j)-m|) ³,ψ∈{c,s}

P(c,s,m)＝|P(c,m)ΘP(s,m)| (6)。

The characteristic pattern generating has 54 width: brightness Fig. 6 width, color characteristic Figure 12 width, direction character Figure 24, object color component feature Figure 12 width.Next utilizing these characteristic patterns to generate visual attention significantly schemes, because each characteristic pattern may comprise the unconspicuous object of conspicuousness, in order to reduce the interference of not obvious object, do not intend all characteristic patterns to be integrated into a significantly figure herein, but adopt the mode that distributes and carry out, first these characteristic patterns being fused to four significantly figure with normalization with across yardstick, is respectively that brightness is significantly schemed color is significantly schemed direction is significantly schemed significantly scheme with object color component again the remarkable figure of this four width is integrated into final significantly figure (S).The object of normalization operation (with N () expression) is that Integral lifting has part strong stimulation peak value in the situation that lacking TD supervision, and entirety suppresses to comprise a large amount of similar peak responses.Its computing method are first image value specification to be turned to [0, M], the difference producing to eliminate amplitude, the secondly mean value of maximal value M and other all local maximums in computed image and then entire image is multiplied by

Across yardstick merge (with " " represent) and process be by each normalized characteristic pattern be compressed to yardstick 4 again corresponding point be added.Brightness is significantly schemed color is significantly schemed direction is significantly schemed significantly scheme with object color component the computing formula that the more remarkable figure of this four width is integrated into final significantly figure (S) is as follows:

$\begin{array}{c}\tilde{I}=\underset{c=2}{\overset{4}{\⊕}}\underset{s=c+3}{\overset{c+4}{\⊕}}N\left(I(c,\tilde{s})\right)\\ \tilde{C}=\underset{c=2}{\overset{4}{\⊕}}\underset{s=c+3}{\overset{c+4}{\⊕}}[N\left(\mathrm{RG}(c,s)\right)+N\left(\mathrm{BY}(c,s)\right)]\\ \tilde{O}=\underset{\mathrm{\θ}=\left\{\mathrm{0,45,90,135}\right\}}{\mathrm{\Σ}}N\left(\underset{c=2}{\overset{4}{\⊕}}\underset{s=c+3}{\overset{c+4}{\⊕}}N\left(O(c,s,\mathrm{\θ})\right)\right)\\ \widetilde{P=}\underset{m=\{m1,m2\}}{\mathrm{\Σ}}N\left(\underset{c=2}{\overset{4}{\⊕}}\underset{s=c+3}{\overset{c+4}{\⊕}}N\left(P(c,s,m)\right)\right)\\ S=(\tilde{I}+\tilde{C}+\tilde{O}+\tilde{P})/4\end{array}---\left(7\right).$

S5 significantly schemes S bilinear interpolation to 0 yardstick, then binaryzation to final notice.

The maximal value of final significantly figure S has defined the position of specific image, namely focus-of-attention.In order to extract ' s focus of attention, the mean value that its bilinear interpolation is calculated to each simply connected region to 0 yardstick (with sizes such as former figure) (for example has five obvious objects to have no more than 5 connected regions in S, corresponding mean value is designated as Si, i ∈ { 1,2,3,4,5}), suppose that two maximum mean values are respectively S ₁, S ₂, and S ₁>S ₂, threshold values T is taken as: T=S ₂+ (S ₁-S ₂)/are 2. following formula binaryzation (f (i, j) represents the gray-scale value of remarkable figure) for figure significantly:

$f(i,j)=\left\{\begin{array}{c}1,f(i,j)>T\\ 0,f(i,j)<T\end{array}\right.---\left(8\right)$

S6 by the final notice of binaryzation significantly scheme S and former input scene graph step-by-step phase with, be partitioned into target object.

A kind of new image partition method based on notice is proposed herein.The method is a kind of multiple dimensioned many characteristic analysis methods, has not only utilized scene information from bottom to top but also has utilized the color characteristic of goal task jointly to form scene image and significantly schemed, thereby then from remarkable figure, having extracted marking area and reach the object of perception.Because the method has added goal task feature, in experiment, also can observe under background or the obvious situation of interfering object, still obtain good effect, this algorithm can meet the utilization demand under most of situations.

Experimental result and the analysis of the image segmentation algorithm of the modified of utilization increase task color character based on notice, the cromogram that the present embodiment image resolution ratio is 520*390.First gather the image of goal task.In order better to extract two kinds of color components of proportion maximum, avoid background interference in image, before input, need background to replace with black (r=g=b=0) or pure white (r=g=b=1), then input system.In system, extract m1, when m2, consider that the image of actual acquisition is due to the reason at illumination and visual angle, even the target of pure white (or black) neither pure white (or black) in image but is concentrated near pure white (or black) place color gamut, therefore ignore pure white and black.Then input visual field Scene image, obtain respectively brightness, color, three passages of direction, 9 yardstick gaussian pyramids (I (δ) by linear filtering, C (δ), O (δ)), in conjunction with m1, m2 obtains object color component gaussian pyramid P (ψ, m), T=0.31 in the present embodiment (effect is better in the time of TT ≈ 80/255).Carry out " center around " and each feature is calculated respectively in normalization operation use across yardstick and merge and the remarkable figure of each passage is calculated in normalization according to each characteristic pattern.Significantly figure fusion obtains final significantly figure again , with after threshold value T binaryzation to its bilinear interpolation to 0 yardstick (with the size such as former figure), with input scene graph phase with.5 visual scene exhibitions that have a representative for Fig. 2, show the operation result of this improvement algorithm, and first row is original image, and second classifies remarkable figure as, and the 3rd row are threshold values binary map after treatment, and last classifies image segmentation result as.

Image segmentation algorithm is realized and is compared

In order to assess and analyze improved algorithm performance herein, we use this algorithm and traditional Itti algorithm, GBVS (Graph-Based Visual Saliency) algorithm to make comparisons.Fig. 3 is the final significantly figure contrast that these several algorithms generate.

Be not entirely accurate perception target by visible this algorithm of upper figure, but be enough for some basic utilizations (such as location, target identification, keeps away barrier etc.).

The remarkable figure generating according to traditional Itti algorithm and GBVS algorithm afterwards can find the marking area in scene image, but the not necessarily target object of object that often this region comprises, especially when target object significance is lower, this phenomenon particularly evident (for example (C1) (D1)).

In order to analyze more accurately and contrast improved algorithm herein, we have defined two variablees and have hit index D and failed index D f.Hit index D relevant with the value of Si (defined) to failed index D f above, suppose that two maximum mean values are respectively S ₁, S ₂, and S ₁>S ₂if, S ₁corresponding object is not to need perceived target (being that the region that significance degree is the highest is not target object), and D=0 represents perception failure, now failed index D f=(S ₂-S ₁)/S ₂; If S ₁corresponding object is to need perceived target (being that the region that significance degree is the highest is target object), D=(S ₁-S ₂)/S ₁, it is better that the larger explanation target of D is significantly hit effect, now failed index D f=0.Table 1 is according to the index D calculating shown in Fig. 3 and failed index D f.By the visible superiority of algorithm herein of data in table

Table 1: this paper method and Itti algorithm and the contrast of GBVS algorithm

Image based on notice cut apart improve algorithm exceed 100 width scene images use in, the overwhelming majority has obtained good result.Exceed 95% result and show that this algorithm can successfully correctly be divided into target object in visual scene.

We have done the perception comparative study of algorithm under white Gaussian noise and salt-pepper noise interference herein simultaneously.Fig. 4 is experimental result, first and second row are respectively the scene graph and the segmentation result that add after white Gaussian noise (average is 0), and third and fourth row are respectively the scene graph and the sensing results that add after salt-pepper noise, and the first to three row noise density equals respectively 0.2,0.6,1.

(accuracy rate represents that pixel that sensing results is correct accounts for the number percent of target complete pixel in the impact of Fig. 4 noise on experimental result; Error rate represents that the pixel of classification error in sensing results accounts for the number percent of target complete pixel).

The foregoing is only the preferred embodiments of the present invention, be not limited to the present invention, obviously, those skilled in the art can carry out various changes and modification and not depart from the spirit and scope of the present invention the present invention.Like this, if these amendments of the present invention and within modification belongs to the scope of the claims in the present invention and equivalent technologies thereof, the present invention is also intended to comprise these changes and modification interior.

Claims (4)

1. the image partition method based on notice TD-BU mechanism, is characterized in that: comprise the following steps:

S1 gathers target image;

S2 uses two to enter gaussian pyramid and produce 9 scalograms the scene image of input and target image;

S3 is according to scene image and target image generating scene characteristic pattern;

S4 significantly schemes according to scene characteristic generating scene, remarkable scene figure is merged to the final notice of generation across yardstick and significantly scheme S;

S5 significantly schemes S bilinear interpolation to 0 yardstick, then binaryzation to final notice;

S6 by the final notice of binaryzation significantly scheme S and former input scene graph step-by-step phase with, be partitioned into target object.

2. image partition method according to claim 1, is characterized in that: described scene characteristic figure comprises color characteristic figure, brightness figure, the direction character figure being extracted by data-driven and the object color component characteristic pattern extracting according to goal task.

3. image partition method according to claim 2, is characterized in that:

Through type (2) calculates brightness figure I (c, s),

I(c,s)＝|I(c)ΘI(s)| (2)

Wherein, yardstick centered by c, c ∈ the pixel in 2,3,4}, s is yardstick around, s=c+d, wherein d ∈ { 3,4};

By r, g, b triple channel expands to R, G, B, tetra-passages of Y, are defined as follows,

$\left\{\begin{array}{c}R=r-(g+b)/2\\ G=g-(r+b)/2\\ B=b-(g+b)/2\\ Y=(r+b)-|r-b|/2-b\end{array}\right.---\left(3\right)$

R, g, b represents respectively three passages of red, green, blue in rgb color model, its brightness is I=(r+g+b)/3, gaussian pyramid R (δ), G (δ), B (δ), Y (δ) is obtained by above four new tunnels respectively;

Through type (4) calculates color characteristic figure,

$\left\{\begin{array}{c}\mathrm{RG}(c,s)=\left|(R\left(c\right)-G\left(c\right))\mathrm{\&Theta;}(G\left(s\right)-R\left(s\right))\right|\\ \mathrm{BY}(c,s)=\left|(B\left(c\right)-Y\left(c\right))\mathrm{\&Theta;}(Y\left(s\right)-B\left(s\right))\right|\end{array}\right.---\left(4\right)$

Wherein, RG (c, s) represents red and green and green and red two colors pair, and BY (c, s) represents blue and yellow and yellow and blue two colors pair;

Through type (5) calculated direction characteristic pattern O (c, s, θ),

O(c,s,θ)＝|O(c,θ)ΘO(s,θ)| (5)

Wherein, θ ∈ { 0 °, 45 °, 90 °, 135 ° } represents four direction,

Calculate by the following method object color component characteristic pattern P (c, s, m),

Computing Principle represents the point in scene image as divided into f (i, j), and f (i, j) corresponding point color value is (r ₀, g ₀, b ₀); The color value that m is corresponding is (r, g, b), and m ∈ (m1, m2): m1, m2 represent respectively two kinds of maximum colors of colour component proportion in target image,

Scene image color is more similar to m, and corresponding point eigenwert is larger;

The too low corresponding point of similarity eigenwert is 0, even | and r ₀-r|or|g ₀– g|or|b ₀– b| is greater than T, and T is threshold value, P (ψ, m)=0; Otherwise

P(ψ，m)＝(1-|f(i,j)-m|) ³,ψ∈{c,s}

P(c,s,m)＝|P(c,m)ΘP(s,m)| (6)。

4. image partition method according to claim 3, is characterized in that: the method that generating scene is significantly schemed S is,

First the scene characteristic figure generating in S3 is fused to four significantly figure with normalization with across yardstick; The remarkable figure of this four width is integrated into final significantly figure S, the described remarkable figure of four width is respectively brightness significantly to scheme again color is significantly schemed direction is significantly schemed significantly scheme with object color component

Normalized computing method: first image value specification is turned to [0, M]; Secondly the mean value of maximal value M and other all local maximums in computed image and then entire image is multiplied by

Merging across yardstick is that each normalized characteristic pattern is compressed to yardstick 4 corresponding point addition again;

The computing formula that final notice is significantly schemed S is as follows:

$\begin{array}{c}\tilde{I}=\underset{c=2}{\overset{4}{\&CirclePlus;}}\underset{s=c+3}{\overset{c+4}{\&CirclePlus;}}N\left(I(c,\tilde{s})\right)\\ \tilde{C}=\underset{c=2}{\overset{4}{\&CirclePlus;}}\underset{s=c+3}{\overset{c+4}{\&CirclePlus;}}[N\left(\mathrm{RG}(c,s)\right)+N\left(\mathrm{BY}(c,s)\right)]\\ \tilde{O}=\underset{\mathrm{\&theta;}=\left\{\mathrm{0,45,90,135}\right\}}{\mathrm{\&Sigma;}}N\left(\underset{c=2}{\overset{4}{\&CirclePlus;}}\underset{s=c+3}{\overset{c+4}{\&CirclePlus;}}N\left(O(c,s,\mathrm{\&theta;})\right)\right)\\ \widetilde{P=}\underset{m=\{m1,m2\}}{\mathrm{\&Sigma;}}N\left(\underset{c=2}{\overset{4}{\&CirclePlus;}}\underset{s=c+3}{\overset{c+4}{\&CirclePlus;}}N\left(P(c,s,m)\right)\right)\\ S=(\tilde{I}+\tilde{C}+\tilde{O}+\tilde{P})/4\end{array}---\left(7\right).$