CN106203430A - A kind of significance object detecting method based on foreground focused degree and background priori - Google Patents

Abstract

A kind of significance object detecting method based on foreground focused degree and background priori, step is as follows: one, Image semantic classification；Two, significance based on foreground focused degree；Three, significance based on background priori；Four, significance optimization fusion；The method obtains concentration class feature by the way of based on Hash coding and concentration class weight, obtains prospect and significantly scheme after being combined with contrast metric based on center priori；By rejecting super-pixel too high with prospect similarity in edge seed, thus obtain background seed, and obtain background significance by the diversity factor calculating each super-pixel and background seed；Finally, build and comprise background item, prospect item and the cost function of smooth item, obtain and optimize final significance by minimizing cost function.The notable uniform highlighting foreground target of figure energy that the present invention obtains, and suppress background noise well, it being widely portable to natural image, beneficially succeeding target detection, Target Segmentation etc. are applied, and have actual application value.

Description

A kind of significance object detecting method based on foreground focused degree and background priori

(1) technical field

The present invention relates to a kind of significance object detecting method based on foreground focused degree and background priori, belong to computer Vision and digital image processing field.The field such as Target Segmentation, target recognition has broad application prospects.

(2) background technology

Human eye is often easier to notice in scene entirely different attribute with surrounding, the obvious object of difference, this Interested in scene can be partially separated out by kind of autosense ability easily.This and the surrounding of object just The significant difference of environment the ability causing human eye to note are referred to as vision significance.Theoretical according to classical significance, vision Attention mechanism can be divided into two classes: top-down attention mechanism and attention mechanism from bottom to top.Top-down attention machine System is by task-driven, and in this mechanism, it is that the consciousness of people and task determine the well-marked target in image for which part, root According to task, having the subjective consciousness of people to set out, the view of people has played the biggest effect in this mechanism.And bottom-up note Meaning mechanism, by data-driven, is namely determined by self-contained information in image.Often by the object in image with Diversity about determines the significance of this object.

At computer vision field, top-down vision noticing mechanism is owing to it is based on the subjective consciousness of people, research Get up the most difficult, so present significance field is with bottom-up vision noticing mechanism for mainly studying discovery.The end of from Upwards classical in the research of mechanism be then that this model is extracted at multiscale space in the Itti model of proposition in 1998 Brightness, color, three, direction feature, as the feature of image, obtain three by filtering and central authorities' difference algorithm around The characteristic pattern of individual feature, finally uses the method linearly adding sum by three kinds of Feature Fusion, obtains final notable figure.Liu et al. exists Within 2007, by well-marked target is carried out feature extraction, obtaining multiscale contrast, center-periphery contrast, color space divides 3 kinds of features of cloth, then be combined with conditional random field models, obtain final significance testing result.Hou et al. was in 2009 Propose a kind of significance computation model based on residual spectrum, at Fourier's domain of variation with the difference of preimage information Yu its redundancy Obtain composing residual risk, change to spatial domain with residual risk contravariant and obtain its Saliency maps.Jiang et al. proposes for 2013 UFO model, by the uniqueness of combining target, centrality and weighs significance like physical property characteristic.Zhu et al. is by calculating district The length that territory profile is connected with border and the ratio of region area, obtained the contour connection feature in region, and basis at this On calculate contrast based on background weight, estimate the significance in each region finally by the method optimized.

Conventional significance model is usual only from target, or only from background.The present invention combines foreground target Feature and the advantage of background priori, it is proposed that the computational methods of a kind of foreground focused degree and background initial point selection method, and profit By the mode optimized, prospect is merged with background, highlight prospect fully and inhibit background.

(3) summary of the invention

(1) purpose of the present invention

In order to make up the deficiency of traditional method, the present invention is from prospect concentration class priori and background priori, it is provided that one Plant significance object detection method based on foreground focused degree and background priori.

Concentration class priori in the present invention combines concentration class and center priori, and background priori is then that the border from image goes out Send out.Because by substantial amounts of image viewing it is found that gathering is compared in the distribution in entire image of the significance target, but Background is then distributed relatively broad, is often distributed among entire image, finds according to this, and the present invention constructs concentration class priori.Separately Outward, according to photography custom, target is normally located in image by paracentral position, is much all by image in existing method Center is as center priori, but this method is easy to mistake occur, in order to solve this problem, the present invention have employed based on The center priori of convex closure, can choose more reliable center priori according to image adaptive.See again by substantial amounts of image Examining it is found that the part of close image boundary mostly typically is background, the most existing a lot of methods then select the border of image As background priori.But reality exists some situations, the border of image comprises a part of significance target, in order to tackle This situation, the present invention proposes the system of selection of a kind of background seed points, thus provides background priori more accurately.

(2) technical scheme

A kind of based on foreground focused degree and background priori the significance object detecting method of the present invention, its concrete grammar walks Rapid as follows:

Step one: Image semantic classification；For subsequent step, first, the gauss hybrid models by structure input picture will Input picture is divided into multilamellar, and utilizes hash conversion to obtain the binary code of each layer；Furthermore, by super-pixel segmentation, input is schemed As being divided into many color similarities, protect the super-pixel on border, and calculate mean place and the average color of each super-pixel；Additionally carry Take the convex closure comprising well-marked target in input picture, using convex closure center as center priori；

Wherein, in " the utilizing hash conversion to obtain the binary code of each layer " described in step one, its practice is as follows: first Build the gauss hybrid models of input picture, represent a kind of color by each composition correspondence of gauss hybrid models, then can will input The color of image is divided into 6 classes, obtains the probability that each pixel belongs to all kinds of simultaneously.Pixel belongs to the probability of each layer can use image Represent, then decompose for 6 parts relative to by input picture, i.e. represent 6 layers of gray level image of degree of membership with gray value；Then This 6 width image is all downsampled to the image that size is 8 × 8, calculates its gray average, gray value is more than the mark of average pixel It is designated as 1, is otherwise 0, thus obtain 64 binary codes that every tomographic image is corresponding；

Step 2: significance based on foreground focused degree；First using the similarity degree between each layer binary code as similar Property estimate, each for the gauss hybrid models of input picture layer is classified, then by calculate all kinds of gatherings based on center priori Degree obtains concentration class feature as weight to carrying out fusion；Calculate each super-pixel again and combine the global contrast of central authorities' priori, To contrast metric.Finally concentration class feature is multiplied with contrast metric, significantly schemes as foreground focused degree；

Wherein, " being classified by each for the gauss hybrid models of input picture layer " described in step 2, its practice is such as Under: first survey using the inverse of the Euclidean distance between binary code corresponding to each tomographic image of gauss hybrid models as similarity Degree, utilizing the clustering method of Alex Rodriguez to be gathered by this 6 tomographic image is 3 classes, respectively the prospect in representative image, background and Dash area, the most each pixel belongs to the probability of this three apoplexy due to endogenous wind K class and is:

$p\left(K\right|I_x)=\underset{k\∈K}{\mathrm{\Σ}}p\left(k\right|I_x)$

Wherein p (k | I_x) it is pixel I_xBelong to the probability of gauss hybrid models kth composition, and this kth becomes to belong to the K class, be equivalent to add several tomographic images belonging to K class and.

Wherein, described in step 2 " again by calculate all kinds of concentration class based on center priori as weight to entering Row fusion obtains concentration class feature ", its process calculated is as follows: add with the three class images that classification is obtained by concentration class for weight With, obtain concentration class characteristic pattern:

$S_C=\underset{K}{\mathrm{\Σ}}p\left(K\right|I_x)*Comp\left(K\right)$

Comp (K) is the concentration class that K class image is corresponding:

$Comp\left(K\right)={\left(\frac{{\mathrm{\Σ}}_{I_x}\left|\right|x-\mathrm{\μ}|{|}^2\·p\left(K\right|I_x)}{{\mathrm{\Σ}}_{I_x}p\left(K\right|I_x)}\right)}^{-1}$

Step 3: significance based on background priori；First the super-pixel being connected with image boundary is obtained as background kind Son；Then, figure binaryzation notable to the prospect obtained in step 2, using the super-pixel that is marked as 1 as foreground seeds point, meter Calculate the similarity degree of other super-pixel and foreground seeds, and determine threshold value；By big with foreground seeds similarity in the super-pixel of border Part super-pixel in threshold value is rejected from background seed, then obtain final background subset；Finally, by calculating each super-pixel With the contrast of background seed, thus obtain background significance；

Wherein, " the calculating the similarity degree of other super-pixel and foreground seeds " described in step 3, its computational methods As follows:

$Sim\left(i\right)=1-\underset{j\∈FS}{\mathrm{\Σ}}\left|\right|S_{fg}\left(i\right)-S_{fg}\left(j\right)\left|\right|$

FS represents foreground seeds point set.

Wherein, " will surpass more than the part of threshold value with foreground seeds similarity in the super-pixel of border described in step 3 Pixel is rejected from background seed ", its process rejected is as follows:

Determined threshold value T of similarity by OSTU algorithm, border super-pixel will be more than threshold value T with foreground seeds similarity Part super-pixel reject from background seed, then obtain final background subset BS；Finally, with each super-pixel and background seed Contrast as to background significance:

$S_{bg}\left(i\right)=\underset{j\∈BS}{\mathrm{\Σ}}\left|\right|c_i-c_j\left|\right|\·(1-||{\mathrm{\μ}}_i-{\mathrm{\μ}}_j|\left|\right)$

Step 4: significance optimization fusion；Fusion problem is considered as optimization problem, builds one and comprise prospect item, background Item and the cost function of smooth item, combine prospect background, obtains final notable figure by minimizing cost function；

In described step 4, first build a cost function, prospect background combined:

Foreground represents prospect item, and Background represents background item, and Smoothness is smooth item；Wherein S (i) For the final significance average of i-th super-pixel, obtain final notable figure by minimizing cost function；α is balance prospect Significance and the background significance weight to final significance power of influence size, λ is the weight of the smooth item effect size of regulation, i.e. Regulate the smoothness of final significance.

Finally by minimizing cost function, obtain final significance S；

By above step, this detection method combines display foreground concentration class and background priori, it is possible to before preferably prominent Scape and suppression background, then can relatively accurately detect image object, for other image processing field such as Target Segmentation, mesh Mark is followed the tracks of has actual application value with target retrieval etc..

(3) compared with prior art, advantages of the present invention:

First, the present invention is using convex closure center as center priori, it is proposed that the meter of the concentration class feature at a kind of relative center Calculation method, and it is combined with global contrast based on center priori, the most complete obvious object can be obtained, and fill Divide the significance of the prospect that highlighted.

Secondly, the present invention proposes a kind of background seed points selection algorithm based on prospect, it is to avoid part is positioned at limit The prospect on boundary is falsely dropped as background seed, thus improves the accuracy of background priori.Significance based on background priori calculates to be had Imitate inhibits the background parts in notable figure.

Finally, the fusion of prospect Yu background significance is considered as optimization problem and processes by the present invention, by building cost function In conjunction with prospect and background, the prospect that takes full advantage of significantly schemes the advantage of figure notable with background, and notable figure is seamlessly transitted, and fills While dividing prominent prospect, inhibit background the most well.

(4) accompanying drawing explanation

Fig. 1 is the FB(flow block) of detection method of the present invention.

(5) detailed description of the invention

In order to be more fully understood that technical scheme, below in conjunction with accompanying drawing, embodiments of the present invention are made further Describe.

The FB(flow block) of the present invention is as it is shown in figure 1, the present invention is a kind of based on foreground focused degree with the significance of background priori Object detecting method, it is as follows that it is embodied as step:

Step one: Image semantic classification

First, build the gauss hybrid models of input picture, represent a kind of face by each composition correspondence of gauss hybrid models Color, then can be divided into 6 classes by the color of input picture, obtains each pixel simultaneously and belongs to the probability of kth class color:

{ω_k,μ_k,∑_kIt is the parameter of gauss hybrid models, pixel belongs to the probability of each layer and can represent with image, then Decompose for 6 parts relative to by input picture, i.e. represent 6 layers of gray level image of degree of membership with gray value.Then by this 6 width figure As being all downsampled to the image that size is 8 × 8, calculate its gray average, gray value is labeled as 1 more than average pixel, no Be then 0, then every tomographic image all can get the binary code of 64.

Then, utilize SLIC algorithm, input picture is too segmented into M=200 super-pixel.And calculate the position of each super-pixel Put μ_iWith color average c_i:

$c_i=\frac{\underset{I_x\∈P_i}{\mathrm{\Σ}}{I}_c}{q_i},{\mathrm{\μ}}_i=\frac{\underset{I_x\∈P_i}{\mathrm{\Σ}}{I}_{\mathrm{\μ}}}{q_i}---\left(2\right)$

Wherein I_cFor pixel I belonged to_xColor vector, I_μFor corresponding space coordinates vector, q_iSuper-pixel block P_iMiddle bag The number of pixels contained.

Finally, the coloured image of input is carried out Harris Corner Detection, the Harris angle point energy of calculating input image Function obtains energy diagram, chooses several points that energy value in energy diagram is maximum, and rejects the point of near image boundaries, obtains calibrated All point of significance are surrounded with a convex closure and represent marking area by true point of significance, and using convex closure center as center first Test.

Step 2: significance based on foreground focused degree

First, using the inverse of the Euclidean distance between binary code corresponding to each tomographic image of gauss hybrid models as similar Degree is estimated, and utilizing the clustering method of Alex Rodriguez to be gathered by this 6 tomographic image is 3 classes, respectively the prospect in representative image, the back of the body Scape and dash area.The most each pixel belongs to the probability of this three apoplexy due to endogenous wind K class:

$p\left(K\right|I_x)=\underset{k\∈K}{\mathrm{\Σ}}p\left(k\right|I_x)---\left(3\right)$

Wherein p (k | I_x) it is pixel I_xBelong to the probability of gauss hybrid models kth composition, and this kth becomes to belong to the K class, be equivalent to add several tomographic images belonging to K class and.With concentration class for weight this three classes image added again and, assembled Degree feature:

$S_C=\underset{K}{\mathrm{\Σ}}p\left(K\right|I_x)*Comp\left(K\right)---\left(4\right)$

Comp (K) is the concentration class that K class image is corresponding, and concrete formula is:

$Comp\left(K\right)={\left(\frac{{\mathrm{\Σ}}_{I_x}\left|\right|x-\mathrm{\μ}|{|}^2\·p\left(K\right|I_x)}{{\mathrm{\Σ}}_{I_x}p\left(K\right|I_x)}\right)}^{-1}---\left(5\right)$

X is pixel I_xCoordinate position, μ is the coordinate position of picture centre.

Then, the computing formula in conjunction with the global contrast of center priori is:

$S_U\left(i\right)=\left(\underset{j\∈M}{\mathrm{\Σ}}\right||c_i-c_j||\·\exp (-\frac{\left|\right|{\mathrm{\μ}}_i-{\mathrm{\μ}}_j\left|\right|}{{{\mathrm{\σ}}_p}^2}\left)\right)\·\exp (-\frac{\left|\right|{\mathrm{\μ}}_i-\mathrm{\μ}\left|\right|}{{{\mathrm{\σ}}_c}^2})---\left(6\right)$

c_iRepresent the color average of super-pixel i, μ_iRepresent the position average of super-pixel i.σ_pFor adjusting color and locus The weight of power of influence, σ_cIt it is then the weight of control centre's priori power of influence.

Finally, with the form of multiplication concentration class feature and contrast metric combined and obtain final prospect significance:

S_fg(i)=S_C(i)·S_U(i) (7)

S_CI () represents the average aggregate degree eigenvalue of super-pixel i.

Step 3: significance based on background priori

First the super-pixel being connected with image boundary is obtained as background seed.Then, to the prospect obtained in step 2 Notable figure binaryzation, using be marked as 1 super-pixel as foreground seeds point, calculate other super-pixel similar to foreground seeds Degree, concrete formula is:

$Sim\left(i\right)=1-\underset{j\∈FS}{\mathrm{\Σ}}\left|\right|S_{fg}\left(i\right)-S_{fg}\left(j\right)\left|\right|---\left(8\right)$

Wherein FS represents foreground seeds point set.

Determined threshold value T of similarity by OSTU algorithm, border super-pixel will be more than threshold value T with foreground seeds similarity Part super-pixel reject from background seed, then obtain final background subset BS.Finally, each super-pixel and background kind are calculated The contrast of son, thus obtain background significance, concrete formula is:

$S_{bg}\left(i\right)=\underset{j\∈BS}{\mathrm{\Σ}}\left|\right|c_i-c_j\left|\right|\·(1-||{\mathrm{\μ}}_i-{\mathrm{\μ}}_j|\left|\right)---\left(9\right)$

Step 4: significance optimization fusion

Build one and comprise prospect item, background item and the cost function of smooth item, prospect background is combined, specifically Formula is:

Foreground represents prospect item, and Background represents background item, and Smoothness is smooth item.Wherein S (i) For the final significance average of i-th super-pixel, obtain final notable figure by minimizing cost function.α is balance prospect Significance and the background significance weight to final significance power of influence size, λ is the weight of the smooth item effect size of regulation, i.e. Regulate the smoothness of final significance.

Finally by minimizing cost function, obtain final significance S.

Claims (6)

1. a significance object detecting method based on foreground focused degree and background priori, it is characterised in that: its concrete grammar Step is as follows:

Step one: Image semantic classification；For subsequent step, first, by building the gauss hybrid models of input picture by input Image is divided into multilamellar, and utilizes hash conversion to obtain the binary code of each layer；Furthermore, by super-pixel segmentation, input picture is divided It is segmented into many color similarities, protects the super-pixel on border, and calculate mean place and the average color of each super-pixel；Additionally extract defeated Enter image comprises the convex closure of well-marked target, using convex closure center as center priori；

Step 2: significance based on foreground focused degree；First survey using the similarity degree between each layer binary code as similarity Degree, classifies each for the gauss hybrid models of input picture layer, then makees by calculating all kinds of concentration class based on center priori Concentration class feature is obtained to carrying out fusion for weight；Calculating each super-pixel again and combine the global contrast of central authorities' priori, it is right to obtain Ratio degree feature；Finally concentration class feature is multiplied with contrast metric, significantly schemes as foreground focused degree；

Step 3: significance based on background priori；First the super-pixel being connected with image boundary is obtained as background seed；So After, figure binaryzation notable to the prospect obtained in step 2, using be marked as 1 super-pixel as foreground seeds point, calculate it The similarity degree of his super-pixel and foreground seeds, and determine threshold value；Border super-pixel will be more than threshold with foreground seeds similarity The part super-pixel of value is rejected from background seed, then obtain final background subset；Finally, by calculating each super-pixel and the back of the body The contrast of scape seed, thus obtain background significance；

Step 4: significance optimization fusion；Fusion problem is considered as optimization problem, build one comprise prospect item, background item and The cost function of smooth item, combines prospect background, obtains final notable figure by minimizing cost function；

In described step 4, first build a cost function, prospect background combined:

Foreground represents prospect item, and Background represents background item, and Smoothness is smooth item；Wherein S (i) is The final significance average of i super-pixel, obtains final notable figure by minimizing cost function；α is that balance prospect is notable Property with the background significance weight to final significance power of influence size, λ is the weight of the smooth item effect size of regulation, i.e. regulates The smoothness of final significance；

Finally by minimizing cost function, obtain final significance S；

By above step, this detection method combines display foreground concentration class and background priori, it is possible to preferably highlight prospect and Suppression background, the most relatively accurately detects image object, for other image processing field such as Target Segmentation, target following and Target retrieval has actual application value.

A kind of significance object detecting method based on foreground focused degree and background priori the most according to claim 1, its It is characterised by: in " the utilizing hash conversion to obtain the binary code of each layer " described in step one, its practice is as follows: first build The gauss hybrid models of input picture, represents a kind of color by each composition correspondence of gauss hybrid models, then by input picture Color is divided into 6 classes, obtains the probability that each pixel belongs to all kinds of simultaneously；Pixel belongs to the probability of each layer and can represent with image, Then decompose for 6 parts relative to by input picture, i.e. represent 6 layers of gray level image of degree of membership with gray value；Then by this 6 width Image is all downsampled to the image that size is 8 × 8, calculates its gray average, and what gray value was more than average pixel is labeled as 1, It is otherwise 0, thus obtains 64 binary codes that every tomographic image is corresponding.

A kind of significance object detecting method based on foreground focused degree and background priori the most according to claim 1, its It is characterised by: " being classified by each for the gauss hybrid models of input picture layer " described in step 2, its practice is as follows: first The inverse of the Euclidean distance between the first binary code corresponding using each tomographic image of gauss hybrid models, as similarity measure, utilizes It is 3 classes that this 6 tomographic image is gathered by the clustering method of Alex Rodriguez, respectively prospect, background and the shadow part in representative image Point, the most each pixel belongs to the probability of this three apoplexy due to endogenous wind K class and is:

$p\left(K\right|I_x)=\underset{k\∈K}{\Σ}p\left(k\right|I_x)$

Wherein p (k | I_x) it is pixel I_xBelong to the probability of gauss hybrid models kth composition, and this kth become to belong to K class, Be equivalent to add several tomographic images belonging to K class and.

A kind of significance object detecting method based on foreground focused degree and background priori the most according to claim 1, its Be characterised by: described in step 2 " again by calculate all kinds of concentration class based on center priori as weight to melting Conjunction obtains concentration class feature ", its process calculated is as follows: add with concentration class for the three class images that classification is obtained by weight and, To concentration class characteristic pattern:

$S_C=\underset{K}{\Σ}p\left(K\right|I_x)*Comp\left(K\right)$

Comp (K) is the concentration class that K class image is corresponding:

$Comp\left(K\right)={\left(\frac{{\mathrm{\Σ}}_{I_x}\left|\right|x-\mathrm{\μ}|{|}^2\·p\left(K\right|I_x)}{{\mathrm{\Σ}}_{I_x}p\left(K\right|I_x)}\right)}^{-1}.$

A kind of significance object detecting method based on foreground focused degree and background priori the most according to claim 1, its Being characterised by: " the calculating the similarity degree of other super-pixel and foreground seeds " described in step 3, its computational methods are as follows:

$Sim\left(i\right)=1-\underset{j\∈FS}{\Σ}\left|\right|S_{fg}\left(i\right)-S_{fg}\left(j\right)\left|\right|$

FS represents foreground seeds point set.

A kind of significance object detecting method based on foreground focused degree and background priori the most according to claim 1, its Be characterised by: described in step 3 " will in the super-pixel of border with foreground seeds similarity more than the part super-pixel of threshold value Reject from background seed ", its process rejected is as follows:

Determined threshold value T of similarity by OSTU algorithm, border super-pixel will be more than the portion of threshold value T with foreground seeds similarity Divide super-pixel to reject from background seed, then obtain final background subset BS；Finally, right with each super-pixel and background seed Than degree as to background significance:

$S_{bg}\left(i\right)=\underset{j\∈BS}{\Σ}\left|\right|c_i-c_j\left|\right|\·(1-||{\mathrm{\μ}}_i-{\mathrm{\μ}}_j|\left|\right).$