CN104463870A - Image salient region detection method - Google Patents

Abstract

The invention provides an image salient region detection method. The method includes the steps that background estimation is conducted on an image generated after super pixel partition is conducted; the contrast ratio of all super pixels and the background is obtained according to the color difference of all the super pixels in the image and the super pixels in the background estimation; a super pixel saliency map is obtained according to the contrast ratio of all the super pixels and the background. By means of the image salient region detection method, robustness can be conducted on background noise of the image and the like, and calculation is easy and fast to conduct.

Description

A kind of detection method for image salient region

Technical field

The present invention relates to technical field of image processing, particularly a kind of detection method for image salient region.

Background technology

In the sensation that the mankind are all, the external information of 70% is had at least to be obtained by vision system.Biological vision system, comprises human visual system, automatically can select and note the position that minority in scene " is correlated with ".For given input picture, as shown in Fig. 1 (a), Fig. 1 (b) shows the mark of its marking area.Human eye can give more concern to the safflower of prospect when observing, and sweeps and mistake greenery and other background area.Biological vision system this in the face of natural scene time the process that can be primarily focused on rapidly on a few significant visual object be called as vision attention select.This ability makes biological tissue that its limited perception cognitive resources is concentrated on maximally related partial data, thus makes them can fast and effeciently process a large amount of signals, survives in the environment of complexity change.

If this mechanism can be introduced art of image analysis, by the marking area that computational resource priority allocation easily causes observer to note to those, greatly will improve the work efficiency of conventional images analytical approach, salient region of image detects and to propose on the basis of this thought just and to grow up.

The region of significant difference both marking area in image is defined by contrasting with neighborhood usually.The modal one of this definition realizes being central authorities-periphery mechanism, and namely central and that periphery difference is large region is marking area.This species diversity can be color distortion, towards difference and texture difference etc.The foremost marking area detection model proposed by Itti and Koch etc. is exactly first carry out multiple dimensioned, multidirectional Gabor convolution to image, extract the color of image, brightness and towards etc. feature, then use difference Gaussian approximation central authorities-periphery poor.In recent research, the people such as Yichen Wei proposes the method based on background distributions priori, carrys out the conspicuousness of estimated image block according to image block to the size of the geodesic distance of image surrounding background.Method based on background priori achieves good effect on some natural images, as shown in Fig. 2 (a) (original image) and Fig. 2 (b) (Saliency maps).But, although the geodesic distance measure used in said method has its rationality, but when processing the image that change of background is larger or texture is very abundant, due to the Accumulation Phenomenon of texture region contrast, the method is estimated the conspicuousness of image can be inaccurate, as shown in Fig. 2 (c) (original image) and 2 (d) (Saliency maps).

Summary of the invention

For problems of the prior art, the invention provides a kind of detection method for image salient region, comprising:

Step 1), background estimating is carried out to the image split through super-pixel;

Step 2), according to the color distortion of the super-pixel in super-pixel each in image and described background estimating, obtain the contrast of each super-pixel and background;

Step 3) obtain super-pixel Saliency maps based on the contrast of each super-pixel and background.

The step 1 of said method) in, the set of the super-pixel of the pixel comprised in described image in the pixel coverage of n, range image border estimated as a setting, wherein n is positive integer.

The step 2 of said method) in, adopt the contrast of each super-pixel and background in following steps computed image:

Step 21), obtain the set of the La*b* color space distance of all super-pixel in this super-pixel and described background estimating according to following formula:

$D_i=\left\{{\left|\right|c_i-c_j\left|\right|}^2\right|\∀S_j\∈\hat{B}\}$

Wherein, D _irepresent super-pixel S _iwith the set of the La*b* color space distance of each super-pixel in background estimating, c _irepresent super-pixel S _ila*b* color, represent background estimating;

Step 22), by the La*b* color space distance in this set by sorting from small to large;

Step 23), using k La*b* color space distance before in this set and as the contrast of this super-pixel and background, wherein k is positive integer.

In said method, step 3) also comprise:

According to the geodesic distance of the super-pixel in super-pixel each in described image and described background estimating, the background obtaining each super-pixel is connective;

For each super-pixel in described image, the contrast of itself and background and background connectedness are carried out linear superposition;

Super-pixel Saliency maps is obtained according to linear superposition result.

In said method, for each super-pixel in described image, using connective as the background of this super-pixel for the minimum value of the geodesic distance of this super-pixel and its color k neighbour in described background estimating, wherein k is positive integer.Wherein, following steps are adopted to calculate the background connectedness of each super-pixel in described image:

Undirected authorized graph is set up for described image, node in this figure comprises super-pixel in described image and virtual background node B, the limit E in this figure and comprises the internal edges that connects neighbouring super pixels and be connected the color k neighbour of super-pixel in described background estimating and the external edge of virtual background node B;

The background calculating each super-pixel according to following formula is connective:

$\mathrm{Connectivity}\left(S_i\right)=\underset{S_1=S_i,S_2,...,S_n=B}{\min }\underset{j=1}{\overset{n-1}{\mathrm{\Σ}}}\mathrm{weight}(S_j,S_{j+1}),\∃(S_j,S_{j+1})\∈E$

Wherein, adjacent two super-pixel S _j, S _j+1weight weight (S _j, S _j+1) for they La*b* color space distance and be expressed as follows:

weight(S _j,S _j+1)＝||c _j-c _j+1|| ²

Wherein, c _jrepresent super-pixel S _jla*b* color, the weight between the super-pixel be connected with virtual background node B and virtual background node B is 0.

In said method, for each super-pixel in described image, according to following formula, the contrast of itself and background and background connectedness are carried out linear superposition:

Saliency(S _i)＝Constrast(S _i)＋α·Connectivty(S _i)

Wherein, Saliency (S _i) represent super-pixel S _iconspicuousness, Constrast (S _i) represent super-pixel S _iwith the contrast of background, Connectivity (S _i) represent super-pixel S _ibackground connective, α is the weight of linear superposition and is be greater than 0 number being less than 1.

Said method can also comprise:

Step 4), super-pixel Saliency maps is processed, obtain pixel significance figure.Wherein, calculate the conspicuousness of each pixel in described image according to following formula, obtain pixel significance figure:

$\mathrm{Saliency}\left(I_p\right)=\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{w}_{\mathrm{pj}}\mathrm{Saliency}\left(S_j\right)$

Wherein, Saliency (I _p) represent pixel I _pconspicuousness, Saliency (S _j) represent super-pixel S _jconspicuousness, weight w _pjbe expressed as follows:

$w_{\mathrm{pj}}=\frac{1}{Z_i}\exp (-\frac{1}{2}(\mathrm{\α}{\left|\right|{\tilde{c}}_p-c_j\left|\right|}^2+\mathrm{\β}{\left|\right|{\tilde{L}}_p-L_j\left|\right|}^2))$

Wherein, and c _jrepresent pixel I respectively _pwith super-pixel S _jla*b* color, and L _jrepresent pixel I respectively _pwith super-pixel S _jcoordinate in described image, α and β represents weight.

In said method, in step 1) also comprise before:

Step 0), texture Fuzzy Processing is carried out to image.

Said method also comprises:

Step 5), binary conversion treatment is carried out to obtained Saliency maps.

Adopt the present invention can reach following beneficial effect:

1, the accuracy detected.Be different from and use the overall situation or local contrast to locate obvious object in the past, the present invention makes full use of the prior imformation of background distributions and to utilize with the contrast of background and connectedness to detect prospect, and detection accuracy has greatly improved.

2, the robustness of tolerance.Can to the robust more of the noise in background estimating by suitably increasing k when calculating background contrasts, and background connectedness can solve part colours a large amount of produced problem in background of foreground object.

Accompanying drawing explanation

Fig. 1 (a) and 1 (b) respectively illustrate an example image and Saliency maps thereof;

Fig. 2 (a) and 2 (b) respectively illustrate another example image and Saliency maps thereof;

Fig. 2 (c) and 2 (d) respectively illustrate another example image and Saliency maps thereof;

Fig. 3 is the process flow diagram of detection method for image salient region according to an embodiment of the invention;

Fig. 4 (a) shows the original image of example;

Fig. 4 (b) be to the original image shown in Fig. 4 (a) carry out texture fuzzy after the texture blurred picture that obtains;

Fig. 4 (c) carries out super-pixel to the texture blurred picture of Fig. 4 (b) to split the result schematic diagram obtained;

Fig. 4 (d) carries out to the image of Fig. 4 (c) result schematic diagram that background estimating obtains;

Fig. 4 (e) is the schematic diagram of the background contrasts of each super-pixel in figure (c);

Fig. 4 (f) is the schematic diagram of the background connectedness of each super-pixel in figure (c);

Fig. 4 (g) is the super-pixel Saliency maps obtained after the background connectedness shown in the background contrasts shown in Fig. 4 (e) and Fig. 4 (f) is carried out linear superposition;

Fig. 4 (h) is the final Saliency maps obtained after the smoothing up-sampling of super-pixel Saliency maps to Fig. 4 (g);

Fig. 5 (a) adopts method provided by the invention and existing methodical PR curve on ASD database;

Fig. 5 (b) adopts method provided by the invention and the evaluation result under existing methodical adaptivenon-uniform sampling on ASD database;

Fig. 6 (a) adopts method provided by the invention and existing methodical PR curve on SED2 database;

Fig. 6 (b) adopts method provided by the invention and the evaluation result under existing methodical adaptivenon-uniform sampling on SED2 database.

Embodiment

Below in conjunction with the drawings and specific embodiments, the present invention is illustrated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, be not intended to limit the present invention.

According to one embodiment of present invention, a kind of detection method for image salient region is provided.

Generally, the method comprises: carry out background estimating to the image split through super-pixel; According to the color distortion of the super-pixel in super-pixel each in image and background estimating, obtain the background contrasts of each super-pixel; Super-pixel Saliency maps is obtained according to the background contrasts of each super-pixel.

Each step of the method is described in detail below in conjunction with Fig. 3.

The first step: texture fuzzy operation is carried out to input picture

In order to remove image (such as, length and width are respectively the input picture I of H, W) human eye such as minor variations in background is insensitive and may form accumulative effect and then affect the high frequency texture change of result of calculation, first can carry out texture Fuzzy Processing to this input picture.Such as, this input picture application structure extracting method is carried out to the suppression of texture region, obtain the image (or claiming texture blurred picture) after texture suppression.The perception that smoothly more can not only meet human eye of this texture part, and the more all even nature of the result that super-pixel subsequently can be made to split.

Fig. 4 (a) shows a width original input picture, and after application structure extracting method carries out texture Fuzzy Processing, effect is as shown in Fig. 4 (b).

It should be noted that, this step not necessarily is necessary, for the image that there is not or have less high frequency texture change, or carried out the input picture of texture Fuzzy Processing, this step can be skipped.

Second step: super-pixel segmentation is carried out to texture blurred picture

After image after obtain texture suppression according to the first step, in order to reduce computation complexity, existing super-pixel partitioning algorithm is then utilized this image to be divided into the super-pixel (i.e. the uniform super-pixel of color) of apparent homogeneity.

Such as, for the texture blurred picture shown in Fig. 4 (b), the result of its super-pixel segmentation is as shown in Fig. 4 (c).

3rd step: utilize super-pixel segmentation result to carry out background estimating according to background priori

Generally, according to the background of background distributions prior estimate image, obtain by the background estimating of super-pixel set expression.

In order to obtain the super-pixel set of the background (i.e. background estimating) estimated, according to the background priori of Saliency maps picture, image surrounding is that the probability of background is very large, therefore in this article, the super-pixel of the pixel comprised in image in the pixel coverage of n, range image border is defined as the background estimating of this image, is shown below:

$\hat{B}=\left\{S_i\right|\min \{x,y,|W-x|,|H-y\left|\right\}\≤n,\∃I_p\∈S_i\}---\left(1\right)$

Wherein, represent background estimating, S _irepresent i-th super-pixel, (x, y) represents pixel I _pcoordinate in the picture, W, H represent width and the height of input picture respectively.N is positive integer, preferably, and n=10.

For the embodiment of Fig. 4 (c), the result of its background estimating is as shown in Fig. 4 (d), and in Fig. 4 (d), the most dark part represents background estimating.

4th step: the background contrasts calculating each super-pixel

As one of ordinary skill in the known, prospect in image usually and background have larger difference, the difference of the background estimating of background and previous step is then relatively little, design the mode of the difference between tolerance super-pixel and image background herein accordingly, i.e. the metric form of the background contrasts of super-pixel.

In one embodiment, the background contrasts (i.e. the contrast of super-pixel and background) of certain super-pixel be calculated, first calculate the La*b* color space distance of each super-pixel in this super-pixel and background estimating according to following formula:

$D_i=\left\{{\left|\right|c_i-c_j\left|\right|}^2\right|\∀S_j\∈\hat{B}\}---\left(2\right)$

Wherein, D _irepresent super-pixel S _iwith the set of the La*b* color space distance of each super-pixel in background estimating, c _irepresent super-pixel S _ila*b* color.

Then, to D _iin distance terms carry out from small to large rearrangement:

${\tilde{D}}_i=<d_{i1},d_{i2},...,d_{\mathrm{iM}}>,d_{i1}\&le;d_{i2}\&le;...\&le;d_{\mathrm{iM}}---\left(3\right)$

Wherein, represent the super-pixel S after sequence _iwith the set of the La*b* color space distance of each super-pixel in background estimating; Each d _im, m=1 ..., M represents that sequence number is the distance terms of m, and wherein M is the number of the super-pixel comprised in background estimating.

For belonging to the super-pixel of background area (abbreviation background), in background estimating, easily find the super-pixel quite similar with its color, therefore its the value of front k item go to zero; And for belonging to the super-pixel of prospect, because the colour-difference exceptional talents of itself and background area make it become obvious object (i.e. prospect), therefore its the value of front k item relatively large.Based on the color distortion of the super-pixel in each super-pixel and background estimating, in one embodiment, the background contrasts of super-pixel is calculated according to following formula:

$\mathrm{Contrast}\left(S_i\right)=\underset{j=1}{\overset{k}{\mathrm{\&Sigma;}}}{d}_{\mathrm{ij}}---\left(4\right)$

Wherein, Constrast (S _i) represent super-pixel S _ibackground contrasts, this background contrasts is super-pixel S _ithe La*b* color space distance sum of k the super-pixel (or claim color k neighbour) the most close with color in background estimating.K is positive integer, preferably, and k=5.

According to the above method calculating background contrasts, each super-pixel in image is calculated to the contrast of itself and background estimating.After the background contrasts obtaining each super-pixel, carry out linear stretch process, thus corresponding super-pixel Saliency maps can be obtained.Such as, Fig. 4 (e) shows the background contrasts of each super-pixel of Fig. 4 (d) with the form of super-pixel Saliency maps, background contrast's angle value of super-pixel is larger, and in Fig. 4 (e), the brightness of this super-pixel is lower.

Above-mentioned detection method for image salient region detects all very effective for the marking area in most of image, and by suitably increasing the value of k, also compares robust for a small amount of noise in background estimating.But when using the method, when the color in prospect occurs in a large number in background estimating, this part prospect may be background by error-detecting, thus the accuracy that impact detects.

For addressing this problem, according to one embodiment of present invention, detection method for image salient region also comprises the steps:

5th step: the background calculating each super-pixel is connective

By finding the observation of foreground object remarkable in image, because the transition in natural image between background is generally relative smooth, relatively large change is then there is to the transition of background by prospect, this change can be understood as the encirclement of object, and super-pixel therefore can be utilized to the geodesic distance of background estimating to measure this change.

In one embodiment, can be connective as the background of super-pixel using the minimum value of the geodesic distance between k the most close for color in super-pixel and background estimating super-pixel.Set forth below is the method calculating this super-pixel background connectedness:

For input picture, set up undirected authorized graph G={V, an E}.Wherein, the node in G is the super-pixel set { S in input picture _iadd a virtual background node B, i.e. V={S _i∪ { B}; Limit in G has two kinds: the internal edges connecting neighbouring super pixels, and connects the color k neighbour of super-pixel in background estimating and the external edge of virtual background node B, i.e. E={ (P _i, P _j) | P _iwith P _jadjacent } ∪ { (P _i, B) | P _ithe color k neighbour of current super-pixel in background estimating }.Preferably, k=5.

By a super-pixel S _igeodesic distance be defined as from S _iset out along the limit weight of the accumulation of shortest path arrival background node B on figure G, then super-pixel S _ithe connective Connectivity (S of background _i) be expressed as follows:

$\mathrm{Connectivity}\left(S_i\right)=\underset{S_1=S_i,S_2,...,S_n=B}{\min }\underset{j=1}{\overset{n-1}{\mathrm{\&Sigma;}}}\mathrm{weight}(S_j,S_{j+1}),\&Exists;(S_j,S_{j+1})\&Element;E---\left(5\right)$

Wherein, adjacent two super-pixel S _j, S _j+1weight weight (S _j, S _j+1) be their La*b* color space distance, i.e. weight (S _j, S _j+1)=|| c _j-c _j+1|| ², c _jrepresent super-pixel S _jla*b* color; Weight between the super-pixel be connected with virtual background node B and background node is 0.

After the background connectedness obtaining each super-pixel, linear stretch can be carried out to the value of this background connectedness, thus represent with figure.

Such as, for the embodiment that Fig. 4 (d) provides, the background that Fig. 4 (f) shows each super-pixel is connective, and wherein color is brighter represents that connective value is larger.

6th step: obtain super-pixel Saliency maps

After the background contrasts calculating each super-pixel and background connectedness, by these two kinds of metric forms are carried out the conspicuousness that linear superposition obtains each super-pixel, thus obtain the Saliency maps of super-pixel according to this conspicuousness, be shown below:

Saliency(S _i)＝Constrast(S _i)＋Connectivity(S _i) (6)

Wherein, Saliency (S _i) represent super-pixel S _iconspicuousness, Constrast (S _i) be the super-pixel S obtained by formula (4) _ibackground contrasts, Connectivity (S _i) the super-pixel S that obtains for formula (5) _ibackground connective.α is the weight of linear superposition and is be greater than 0 number being less than 1, preferably, and α ∈ [0.3,0.6].

For background contrasts and background interconnectedness shown in Fig. 4 (e) and 4 (f), the result after two kinds of metric form linear superposition is as shown in Fig. 4 (g).

7th step: super-pixel Saliency maps is processed and obtains final pixel significance figure

Because the Saliency maps obtained after above-mentioned superposition is still base unit with super-pixel, obtain the more accurate Saliency maps represented by pixel, aftertreatment can be done to super-pixel Saliency maps according to the color of pixel in former figure and positional information, such as perform the conspicuousness that the smooth operation being similar to up-sampling obtains each pixel, be shown below:

$\mathrm{Saliency}\left(I_p\right)=\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{w}_{\mathrm{pj}}\mathrm{Saliency}\left(S_j\right)---\left(7\right)$

Wherein, Saliency (I _p) represent pixel I _pconspicuousness, Saliency (S _j) represent super-pixel S _jconspicuousness, level and smooth weight w _pjcalculating is expressed as follows:

$w_{\mathrm{pj}}=\frac{1}{Z_i}\exp (-\frac{1}{2}(\mathrm{\&alpha;}{\left|\right|{\tilde{c}}_p-c_j\left|\right|}^2+\mathrm{\&beta;}{\left|\right|{\tilde{L}}_p-L_j\left|\right|}^2))---\left(8\right)$

Wherein, and c _jbe respectively pixel I _pwith super-pixel S _jla*b* color vector, and L _jbe respectively pixel I _pwith super-pixel S _jcoordinate vector in the picture, α and β represents the weight to color and position respectively, preferably, α=1/30, β=1/30.

For the super-pixel Saliency maps of Fig. 4 (g), the final pixel significance figure that Fig. 4 (h) obtains after giving and carrying out aftertreatment to it.

In a further embodiment, further process can also be done according to embody rule to obtained Saliency maps, such as, carry out binaryzation, thus the marking area more in saliency maps picture.

Said method according to human visual system to the high-frequency information minor variations of background texture (in the such as image) the insensitive fact in image, utilize the priori of background distributions in image, for the visual characteristic of obvious object in conjunction with the marking area in background contrasts and background detection of connectivity image, it can effectively detect obvious object and reduce Accumulation Phenomenon.

For verifying the validity of detection method for image salient region provided by the invention, inventor tests on ASD database and SED2 database by the appraisal procedure that marking area test problems is usual, experiment adopts following two evaluation metricses: precision-recall curve and Saliency maps carry out the precision (precision) of adaptive threshold fuzziness result, the contrast of recall rate (recall) and F value (F-measure).

Experiment on ASD database and evaluation result as follows:

ASD database is a subset of MSRA database, contains 1000 width test patterns, often opens image and marks there being the conspicuousness of artificial Pixel-level.On ASD database, the salient region detecting method after the background contrasts propose the present invention and background connectedness merge and 11 kinds of current best salient region detecting methods contrast.These 6 kinds of methods are respectively: FT method, RC method, GS method, SF method, GC method, MC method, method provided by the invention is abbreviated as TB method.The evaluation result of the Saliency maps that various method produces on ASD database is as shown in Fig. 5 (a) He 5 (b).As seen from the figure, detection method provided by the invention reaches the result comparable with current method.

Experiment on SED2 database and evaluation result as follows:

SED2 database is the database that every width image comprises two obvious objects.Comprise 100 width images in database, and provide the foreground pixel of corresponding accurately artificial mark.On SED2 database, detection method provided by the invention and 6 kinds of current good salient region detecting methods are contrasted.These 10 kinds of methods are respectively: FT method, RC method, SF method, GC method, MC method, method provided by the invention is abbreviated as TB method.The evaluation result of the Saliency maps that various method produces on SED2 database is as shown in Fig. 6 (a) He 6 (b).As seen from the figure, method provided by the invention achieves experimental result best on this database, and the F-measure index best result compared in existing method wherein after adaptive threshold fuzziness improves 3.6 percentage points.It should be noted that and understand, when not departing from the spirit and scope of the present invention required by accompanying claim, various amendment and improvement can be made to the present invention of foregoing detailed description.Therefore, the scope of claimed technical scheme is not by the restriction of given any specific exemplary teachings.

Claims (11)

1. a detection method for image salient region, comprising:

Step 1), background estimating is carried out to the image split through super-pixel;

Step 2), according to the color distortion of the super-pixel in super-pixel each in image and described background estimating, obtain the contrast of each super-pixel and background;

Step 3) obtain super-pixel Saliency maps based on the contrast of each super-pixel and background.

2. method according to claim 1, wherein, in step 1) in, by described image, the set comprising the super-pixel of the pixel in the pixel coverage of n, range image border is estimated as a setting, and wherein n is positive integer.

3. method according to claim 1 and 2, wherein, in step 2) in, adopt the contrast of each super-pixel and background in following steps computed image:

Step 21), obtain the set of the La*b* color space distance of all super-pixel in this super-pixel and described background estimating according to following formula:

$D_i=\left\{{\left|\right|c_i-c_j\left|\right|}^2\right|\&ForAll;S_j\&Element;\hat{B}\}$

Wherein, D _irepresent super-pixel S _iwith the set of the La*b* color space distance of each super-pixel in background estimating, c _irepresent super-pixel S _ila*b* color, represent background estimating;

Step 22), by the La*b* color space distance in this set by sorting from small to large;

Step 23), using k La*b* color space distance before in this set and as the contrast of this super-pixel and background, wherein k is positive integer.

4. method according to claim 1 and 2, wherein, step 3) also comprise:

According to the geodesic distance of the super-pixel in super-pixel each in described image and described background estimating, the background obtaining each super-pixel is connective;

For each super-pixel in described image, the contrast of itself and background and background connectedness are carried out linear superposition;

Super-pixel Saliency maps is obtained according to linear superposition result.

5. method according to claim 4, wherein, for each super-pixel in described image, using connective as the background of this super-pixel for the minimum value of the geodesic distance of this super-pixel and its color k neighbour in described background estimating, wherein k is positive integer.

6. method according to claim 5, wherein, adopts following steps to calculate the background connectedness of each super-pixel in described image:

Undirected authorized graph is set up for described image, node in this figure comprises super-pixel in described image and virtual background node B, the limit E in this figure and comprises the internal edges that connects neighbouring super pixels and be connected the color k neighbour of super-pixel in described background estimating and the external edge of virtual background node B;

The background calculating each super-pixel according to following formula is connective:

$\mathrm{Connectivity}\left(S_i\right)=\underset{S_1=S_1,S_2,...,S_n=B}{\min }\underset{j=1}{\overset{n-1}{\mathrm{\&Sigma;}}}\mathrm{weigth}(S_j,S_{j+1}),\&Exists;(S_j,S_{j+1})\&Element;E$

Wherein, adjacent two super-pixel S _j, S _j+1weight weight (S _j, S _j+1) for they La*b* color space distance and be expressed as follows:

weight(S _j,S _j+1)＝||c _j-c _j+1|| ²

Wherein, c _jrepresent super-pixel S _jla*b* color, the weight between the super-pixel be connected with virtual background node B and virtual background node B is 0.

7. method according to claim 4, wherein, for each super-pixel in described image, according to following formula, the contrast of itself and background and background connectedness are carried out linear superposition:

Saliency(S _i)＝Constrast(S _i)+α·Connectivity(S _i)

Wherein, Saliency (S _i) represent super-pixel S _iconspicuousness, Constrast (S _i) represent super-pixel S _iwith the contrast of background, Connectivity (S _i) represent super-pixel S _ibackground connective, α is the weight of linear superposition and is be greater than 0 number being less than 1.

8. method according to claim 1 and 2, also comprises:

Step 4), super-pixel Saliency maps is processed, obtain pixel significance figure.

9. method according to claim 8, wherein, calculates the conspicuousness of each pixel in described image, obtains pixel significance figure according to following formula:

$\mathrm{Saliency}\left(I_p\right)=\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{w}_{\mathrm{pj}}\mathrm{Saliency}\left(S_j\right)$

Wherein, Saliency (I _p) represent pixel I _pconspicuousness, Saliency (S _j) represent super-pixel S _jconspicuousness, weight w _pjbe expressed as follows:

$w_{\mathrm{pj}}=\frac{1}{Z_i}\exp (-\frac{1}{2}(\mathrm{\&alpha;}{\left|\right|{\tilde{c}}_p-c_j\left|\right|}^2+\mathrm{\&beta;}{\left|\right|{\tilde{L}}_p-L_j\left|\right|}^2))$

Wherein, and c _jrepresent pixel I respectively _pwith super-pixel S _jla*b* color, and L _jrepresent pixel I respectively _pwith super-pixel S _jcoordinate in described image, α and β represents weight.

10. method according to claim 1 and 2, wherein, in step 1) also comprise before:

Step 0), texture Fuzzy Processing is carried out to image.

11. methods according to claim 8, also comprise:

Step 5), binary conversion treatment is carried out to obtained Saliency maps.