CN103208095B - A kind of quick slit cropping method based on band and neighbourship degree - Google Patents

Abstract

A kind of quick slit cropping method based on band and neighbourship degree belongs to image field. It is that quick slit cropping method is improved, and uses restraint in conjunction with the thought of neighbourship degree and band, makes finedraw can more be evenly distributed in image regional, and then obtains more gratifying scaled results. First the method divides some uniformly-spaced bands image, then utilize Saliency maps that this image is corresponding to calculate the average importance of each band, obtain the target size of each band according to the target size of the importance of band and entire image by solving the optimization problem of a Problem with Some Constrained Conditions, finally in each band, adopt and remove to obtain target size image in conjunction with the quick slit cropping method of neighbourship degree. This method will be far superior to other methods based on slit cropping in computational speed, and the resultant image quality simultaneously obtaining is better than quick slit cropping method, therefore, has certain using value and meaning.

Description

A kind of quick slit cropping method based on band and neighbourship degree

Technical field

The present invention relates to the image scaling technology of content-based perception in technical field of image processing, concreteRelate to a kind of research and realization of the quick slit cropping method based on band and neighbourship degree.

Background technology

Along with the fast development of Internet technology and improving constantly of digital device level of hardware, be suitable forConstantly update and regenerate in the terminal device of heterogeneous networks, for example TV, notebook computer, PDA, handMachine etc., they have dissimilar display size, but most of image is all with a certain fixed rulerVery little and the ratio of width to height is made, and while showing them, conventionally needs when getting at dissimilar digital deviceSize and the ratio of width to height of changing them, adapt to various display devices. Traditional image-scaling methodMainly be divided into 3 kinds, wherein a kind of original image obtained to target size by interpolation or down-samplingImage, so-called even convergent-divergent Scaling technology that Here it is, but when changing aspect ratioWhile carrying out even convergent-divergent, it tends to cause the stretcher strain of image main contents. Another kind method justBe simply to cut Cropping technology, it obtains target size by dismissing image border content,This can bring the loss compared with multiple-content information. The third method is exactly to select picture traverse or heightLittle scaling as a whole scaling carries out uniform zoom to original image, this way also byBe called " mailbox " letterboxing technology, its problem is often can be at equipment up and down or left and right twoLimit produces dark border, can not utilize fully the limited source of screen of mobile device. Above-mentioned 3 kinds of biographiesThe Zoom method of system is not all considered the content of image. In order to make the user can be more easypro on terminal deviceWatch image suitablely, need to carry out convergent-divergent processing to it according to the significance level of picture material, come adaptiveShould be in the terminal presentation facility of different network type, different size, different proportion. So content-basedHow the image scaling technology of perception is better presented at diversified numeral by image with solving exactlyProblem on equipment.

After 2007, the image scaling technology of content-based perception obtains very large development, can be by threeKind of approach changes the size of image: according to the importance degree deletion of image pixel or copy pixel, polymerizationOr expansion pixel and two kinds of modes that method combines, method therefore can roughly be divided three classes it.One class is the discrete method taking slit cropping SeamCarving as representative, and another kind of is to reflect with non-homogeneousPenetrate the continuation method that Warping is representative, also having a kind of is exactly the method that multiple operation combines. BaseBe exactly logical processing in image in unnoticed pixel in the advantage of the zoom technology maximum of perception of contentHold the size that changes image, protect to greatest extent important area that deformation does not occur, as far as possible simultaneouslyThe global information of reservation original image.

The discrete method of reducing based on finedraw is the discrete operations of image being carried out to Pixel-level, heavy in reservationWhen wanting region, constantly delete or copy unessential pixel finedraw and change the size of image. ItsFeature is to keep to greatest extent the important area in image, removes minimum energy region. ButAfter the inessential information in image is removed to a certain degree, continuing employing seam cutting method will certainlyBring loss and the object distortion of important information, thereby cause visual quality of images fast-descending. AlthoughThe discreteness of such algorithm more can embody the flexibility of processing procedure, but also more easily destroys in imageStructural information and also can produce crenellated phenomena and obtain rough result images. And due at every turnAfter having calculated cumlative energy cost matrix, all only process a pixel finedraw, need the iterative computation repeatingGo out some finedraws and could realize the convergent-divergent of image, this will directly cause such algorithm ubiquity timeThe problem that computation complexity is higher.

Non-homogeneous mapping method main thought is to set up the optimum mapping of source images to target image, mappingIn process, add some constraintss to protect picture material. First automatically determine each according to picture materialThe importance of pixel, then carries out non-homogeneous convergent-divergent according to the importance of pixel to image, important districtSimilarity transformation is taked in territory substantially, and nonlinear transformation is taked in inessential region, makes image importantRegional deformation and scaling are less, and deformation is diffused in to non-important area. With with non-homogeneous mappingWarping is that the continuation method majority of representative is all by solving the optimization problem with Prescribed PropertiesObtain optimum distortion of the mesh, the deformation of constraint important area grid is less and distortion of the mesh is disperseedIn inessential region, finally utilize mapping technology to obtain result images. because the method is according to figureAdopt different convergent-divergent strategies as the importance degree of content, utilize non-important area to carry out " hiding " deformation,But the excessive distortion of non-important area may change the semantic information of image. In addition, if imageIn do not have enough non-important areas to hide deformation, continuation method may cause the tight of important contentHeavily distortion. Because the essence of continuous method is equivalent to image to carry out resampling, such operation gestureMust cause result images to occur fuzzy effect.

Consider the complementary characteristic that distinct methods is intrinsic, some researchers propose several method to tie mutuallyIncompatible image is carried out to convergent-divergent processing, be called for short mixed method. The key of this problem is that how to confirm is variousThe amount of method of operating and operating sequence thereof. But determining after the quantity and order of multioperation method,Its difficult point be how to design a kind of effectively evaluating tolerance mechanism weigh objective result image withThe similarity of source images. This interpretational criteria is in fact also the institute of the main difference of different mixing modes. Because such algorithm synthesis the advantage of several different methods, but simultaneously it also introduces lacking of the whole bag of tricksPoint, so how optimum combination the whole bag of tricks determines best operating sequence and operational ton, needsA kind of effective quality evaluating method removes to calculate source images and decides different contractings from the similarity of target imagePut the switching point between method. But the time complexity of multioperation method is higher mostly, so howThe processing speed of going to accelerate multioperation method is also the problem that need to think better of.

Summary of the invention

The invention provides a kind of rapid image convergent-divergent that is applicable to different user terminals of content-based perceptionTechnology, can be presented on dissimilar terminal device the image of fixed dimension adaptively, withTime can keep as much as possible important content in original image, keep key object undistorted, keep figureImportant structure in picture is that spatial relation is relatively constant, to ensure that optimized image views and admires quality.

Because current image-scaling method has pluses and minuses and limitation separately, the present invention is directed to finedrawThe slow-footed problem of method of cutting out, and occur breaking because the concentrated removal of finedraw causes result imagesLayer or discontinuous problem, provide a kind of image scaling technology based on band and neighbourship degree fast.

In order to realize the problems referred to above, the invention provides a kind of effectively fast based on band and neighbourship degreeSpeed slit cropping image-scaling method. Without loss of generality, to change picture altitude as example (change imageThe method of width is similar with it), the method specifically comprises:

1) input one width size is the original image of W*H, and to set its target size be W*H_T，WFor picture traverse, the height that H is original image, H_TFor the height of target image;

2) original image is divided into the equidistant horizontal strip Strips of N bar;

3) calculate the importance S of each band according to Saliency maps Saliency corresponding to former figure_i；

4) go to calculate each according to the height H T of the importance values Si of each band and target imageThe object height hi ' of band. In original image, each strip width is h=H/N, eachThe target width h that band is corresponding_i'＝{h₁',h₂'...h_N'}^T, i=1,2 ... N is by solving optimizationProblem obtains, in this optimization problem, F (hi ') be its object function, define orderThe deformation quantity of mark band, formula (2) is its constraints:

$F\left({h_i}^{\′}\right)=\min \underset{i=1}{\overset{N}{\mathrm{\Σ}}}{S}_i{(h-{h_i}^{\′})}^2---\left(1\right)$

$\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{h_i}^{\′}=H_T---\left(2\right)$

(1) in formula, S_iRepresent i the importance values that band is corresponding in original image, h and h_i' generation respectivelyOriginal height and the object height of a table the i horizontal strip. (2) H in formula_TFor the target of image highDegree, N represents band total quantity.

5), according to the original height h of each band and object height hi ', can determine each band needsFinedraw quantity the Numi=hi '-h removing. Then in each band according to neighbourship degree andCumulative energy is searched for finedraw to be processed. Obtain target size by deleting these finedrawsImage, as follows in the concrete steps of each band search finedraw:

1. the Optimum Matching relation between calculating pixel:

Adopt the quick slit cropping method of the yellow China of prior art proposition by maximizing between adjacent columnsThe weight of edges matched and set up the Optimum Matching relation between pixel in original image between pixel. OftenBetween two row, carry out iterative computation and determine the Optimum Matching relation of pixel between row, and then go to obtain wholeWith the Optimum Matching relational matrix AR of Strip_W×h。

2. calculate the neighborhood relationships between finedraw seams according to pixel Optimum Matching relation:

The horizontal strip that is h for a height, we need to search for h bar finedraw. Definition finedraw existsK is listed as its finedraw label of pixel element E that m is capable^k(m) mark. In addition, we define accumulationNeighborhood relationships matrix A N_h×h, this matrix according to two finedraw elements of each row in whether phase of important areaNeighbour accumulates renewal. Important area is by the remarkable figure saliencymap of the image work of averagingCarrying out binaryzation for threshold value obtains again. Set up the process of finedraw neighborhood relationships matrix A N from right to leftCarry out, on the right side, one row are k=W, and the finedraw label parameter of pixel is initialized as:

E^W(m)＝m,m＝1,2,...h(3)

Neighborhood relationships matrix A N is initialized as:

AN(x,y)＝0,x＝1,2,...hy＝1,2,...h(4)

Further according to the Optimum Matching relational matrix AR between pixel according to following Policy Updates finedraw markNumber parameters E (m) and neighborhood relationships matrix A N:

E^k(m+AR(m,k+1))＝E^k+1(m),m＝1,2,...h,k＝1,2,...W(5)

$\left\{\begin{array}{c}{\mathrm{AN}}^p(E^k\left(m\right),E^k(m-1))={\mathrm{AN}}^{p-1}(E^k\left(m\right),E^k(m-1))+1,k=W-1,...1,m=2,...h\\ {\mathrm{AN}}^p(E^k(m-1),E^k\left(m\right))={\mathrm{AN}}^{p-1}(E^k(m-1),E^k\left(m\right))+1,k=W-1,...1,m=2,...h\\ {\mathrm{AN}}^p(E^k\left(m\right),E^k(m+1))={\mathrm{AN}}^{p-1}(E^k\left(m\right),E^k(m+1))+1,k=W-1,...1,m=1,...h-1\\ {\mathrm{AN}}^p(E^k(m+1),E^k\left(m\right))={\mathrm{AN}}^{p-1}(E^k(m+1),E^k\left(m\right))+1,k=W-1,...1,m=1,...h-1\\ p=(W-k)\×h+m\end{array}\right.---\left(6\right)$

In formula (5), AR (m, k) is the value at the capable k column position of m place in Optimum Matching relational matrix AR,E^k(m) be the finedraw number value of the capable k row of m pixel. In formula (6), AN^pRepresent to have upgraded pThe value of neighborhood relationships matrix A N when pixel, p=1 represents the pixel of band right column the first row. UntilBe updated to last pixel element of first row, we can obtain final neighborhood relationships matrixAN_h×h。

3. calculate the neighbourship degree Neighborability between finedraw:

Neighbourship degree is used for representing adjacent probability between two finedraws, and it will be used for removing after certain finedraw,Its adjacent finedraw is carried out to energy weighting. For example, to the m article of finedraw of removing, we calculate itNeighbourship degree sum is:

$T\_\mathrm{AN}\left(m\right)=\underset{n=1}{\overset{h}{\mathrm{\Σ}}}\mathrm{AN}(m,n)---\left(7\right)$

Then, calculate respectively other finedraws and the m article of probability that finedraw is adjacent. For example, n article thinThe probability calculation that seam and m article of finedraw are adjacent is as follows:

$\mathrm{Neighborability}(m,n)=\frac{\mathrm{AN}(m,n)}{T\_\mathrm{AN}\left(m\right)},n=\mathrm{1,2},...h,m=\mathrm{1,2},...h---\left(8\right)$

4. cumlative energy and the neighbourship degree thereof of comprehensive finedraw are searched for finedraw to be processed:

A. remove the finedraw m of cumlative energy minimum according to least energy principle.

B. then use formula (9) to upgrade its cumlative energy to remaining finedraw,

$\stackrel{\‾}{\mathrm{AE}\left(n\right)}=\mathrm{AE}\left(n\right)+\mathrm{AE}\left(n\right)*w\left(n\right),n=\mathrm{1,2},...h---\left(9\right)$

In formula (9), AE (n) represents the cumlative energy of n article of finedraw, and w (n) calculates and gets according to neighbourship degree,Specific definition is:

w(n)＝C*Neighborability(m,n),n＝1,2...h(10)

In formula (10), C is a constant, and in the present invention, value is 1, is used for regulating neighbourship degree to finedrawThe influence degree distributing. Neighborability (m, n) represents the phase of n article of finedraw and m article of finedrawNeighbour's degree.

C. repeating step A and B, we can find the Num that all will remove_i=h-h_iBar is optimum thinSeam, removes corresponding finedraw, obtains the horizontal strip strip of object height.

It is that quick slit cropping method is improved, in conjunction with the thought of neighbourship degree and band in addition approximatelyBundle, makes seam can more be evenly distributed in image regional, and then obtains more gratifyingResult images. First the present invention divides some uniformly-spaced bands image, then utilizes this image correspondenceSaliency maps calculate the average importance of each band, according to the importance of band and entire imageTarget size obtains the target chi of each band by solving the optimization problem of a Problem with Some Constrained ConditionsVery little, finally in each band, adopt and remove to obtain target chi in conjunction with the quick slit cropping method of neighbourship degreeVery little image. The present invention will be far superior to other methods based on slit cropping in computational speed, simultaneouslyThe resultant image quality obtaining is better than quick slit cropping method, and therefore, the present invention has certainUsing value and meaning.

Brief description of the drawings:

Fig. 1 is the original image that this example size to be processed is 500*330.

Fig. 2 is the schematic diagram that original image is divided into 5 equidistant horizontal strips.

Fig. 3 is the Saliency maps that original image is corresponding.

Fig. 4 sets up pixel level Optimum Matching to be related to schematic diagram.

Fig. 5 is finedraw labelled notation the 1st step schematic diagram.

Fig. 6 is finedraw labelled notation the 2nd step schematic diagram.

Fig. 7 is finedraw labelled notation the 3rd step schematic diagram.

Fig. 8 is finedraw labelled notation the 4th step schematic diagram.

Fig. 9 is finedraw labelled notation the 5th step schematic diagram.

Figure 10 is finedraw labelled notation the 6th step schematic diagram.

Figure 11 is the finedraw distribution map that does not adopt neighbourship degree method.

Figure 12 is the scaled results figure that adopts neighbourship degree method.

Figure 13 is the scaled results figure that does not adopt neighbourship degree method.

Figure 14 is the convergent-divergent front and back comparison diagram that adopts neighbourship degree method.

Figure 15 is the convergent-divergent front and back comparison diagram that does not adopt neighbourship degree method.

Detailed description of the invention:

1) input one width size is the original image (Fig. 1) of 500*330, and sets its target size and be500*200；

2) original image is divided into 5 equidistant horizontal strips (Fig. 2);

3) Saliency that obtains source images schemes (Fig. 3), then calculates each according to Saliency figureThe importance values of band, i.e. the mean value of all pixel significance Saliency in each band,Computing formula is as follows:

$S_i=\frac{1}{(y_i-y_{i-1})\&times;W}\mathrm{\&Sigma;S}(x,y),y_{i-1}\&le;y<y_i,0\&le;x<W,1\&le;i\&le;N---\left(1\right)$

In formula (1), Si is i the importance values that horizontal strip is corresponding, y_i-1And y_iRepresent respectively iThe up-and-down boundary position of individual band, the width that W is image, s (x, y) is the conspicuousness value of pixel (x, y).

4) according to the importance values S of each band_iAnd the height H of target image_T=200 go to calculate oftenThe object height h of individual band_i'. In original image, each strip width isH=H/N=500/5, the target width h that each band is corresponding_i'＝{h₁',h₂'...h_N'}^T,i＝1,2,...NCan obtain by solving optimization problem, in this optimization problem, F (h_i') be itsObject function, has defined the deformation quantity of target stripe, and formula (3) is its constraints:

$F\left({h_i}^{\&prime;}\right)=\min \underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{S}_i{(h-{h_i}^{\&prime;})}^2---\left(2\right)$

$\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{h_i}^{\&prime;}=H_T---\left(3\right)$

(2) in formula, S_iRepresent i the importance values that band is corresponding in original image, h and h_i' generation respectivelyOriginal height and the object height of a table the i horizontal strip. (3) H in formula_TFor the target of image highDegree degree, N represents band total quantity.

5) according to the original height h=100 of each band and object height h_i', we can determine eachThe finedraw quantity Num that band need to be removed_i=h_i'-h. Then basis in each bandNeighbourship degree and cumulative energy are searched for finedraw to be processed. Obtain by deleting these finedrawsTarget size image, as follows in the concrete steps of each band search finedraw:

1. the Optimum Matching relation between calculating pixel:

The quick slit cropping method that adopts the yellow China of prior art to propose is set up original imageOptimum Matching relation between middle pixel. The method by maximizing between adjacent columnsBetween pixel the weight of edges matched and, utilize formula (4) to set up the optimum between pixelMatching relationship (taking the matching relationship of setting up pixel between row as example):

F(m)＝max{F(m-1)+w(m,m),F(m-2)+w(m,m-1)+w(m-1,m)},(4)

m＝1,2,...h

Wherein, F (m) is illustrated between k row and k+1 row pixel front m to pixelExcellent edges matched weight sum, w (m, n) represents pixel I (k, m) and pixel I (k+1, n)Between edges matched weighted value.

Edge weights value w (m, n) is defined as:

$w(m,n)=\left\{\begin{array}{cc}A(k,m)\&CenterDot;M(k+1,n),& |m-n|\&le;1\\ -\&infin;,& \mathrm{otherwise}\end{array}\right.---\left(5\right)$

Wherein, A (k, m) representative is listed as the pixel obtaining according to matching relationship from first to kThe cumlative energy of I (k, m), M (n, k+1) representative is listed as k+1 and is listed as according to consequent from endThe cumlative energy of the pixel I (k+1, n) that ENERGY METHOD obtains.

Determine the optimum of pixel between row by the computing formula (4) of iteration between every two rowMatching relationship, and then remove to obtain the Optimum Matching relational matrix of whole strips S tripAR_W×H, between row, the matching relationship of pixel only has 3 kinds of situations (Fig. 4): upper left, a left sideAnd lower-left, at matrix A R_W×HElement value corresponding to middle difference is-1,0,1.

2. calculate the neighbouring relations between finedraw seams according to pixel Optimum Matching relation:The vertical band strip that is h=100 for a height, we need to search for 100Bar finedraw. Definition finedraw is listed as at k its finedraw label for pixel element that m is capable E^k(m) mark. In addition, definition accumulation neighborhood relationships matrix A N_h×h, this matrix rootAccumulate renewal according to two finedraw elements of each row whether important area is adjacent,The matrix size 100 × 10 of AN. Important area is by the remarkable figure of imageSaliencymap averages and carries out binaryzation as threshold value again and obtain. Set up finedrawThe process of neighborhood relationships matrix is upgraded from right to left, at the right side one rowK=W=500, the finedraw label parameter of pixel is initialized as:

E^W(m)＝m,m＝1,2,...h(6)

Neighborhood relationships matrix A N is initialized as:

AN(x,y)＝0,x＝1,2,...hy＝1,2,...h(7)

Further according to the Optimum Matching relation between pixel according to following Policy Updates finedraw markNumber parameters E (m) and neighborhood relationships matrix A N:

E^k(m+AR(m,k+1))＝E^k+1(m),m＝1,2,...h,k＝1,2,...W(8)

$\left\{\begin{array}{c}{\mathrm{AN}}^p(E^k\left(m\right),E^k(m-1))={\mathrm{AN}}^{p-1}(E^k\left(m\right),E^k(m-1))+1,k=W-1,...1,m=2,...h\\ {\mathrm{AN}}^p(E^k(m-1),E^k\left(m\right))={\mathrm{AN}}^{p-1}(E^k(m-1),E^k\left(m\right))+1,k=W-1,...1,m=2,...h\\ {\mathrm{AN}}^p(E^k\left(m\right),E^k(m+1))={\mathrm{AN}}^{p-1}(E^k\left(m\right),E^k(m+1))+1,k=W-1,...1,m=1,...h-1\\ {\mathrm{AN}}^p(E^k(m+1),E^k\left(m\right))={\mathrm{AN}}^{p-1}(E^k(m+1),E^k\left(m\right))+1,k=W-1,...1,m=1,...h-1\\ p=(W-k)\&times;h+m,p=\mathrm{1,2},...W\&times;h\end{array}\right.---\left(9\right)$

In formula (8), AR (m, k) is the value at the capable k column position of m place in Optimum Matching relational matrix,E^k(m) be the finedraw number value of the capable k row of m pixel, h represents in original image thisBand height, the width of W presentation graphs picture. In formula (9), AN^pRepresent to have upgraded pThe value of neighborhood relationships matrix A N when individual pixel, this band right column the first rowPixel be updated in an orderly manner last pixel element of first row, we can obtainTo final neighborhood relationships matrix A N_h×h。

At length set forth the foundation of neighborhood relationships matrix A N below with a simple exampleProcess:

A. the width of supposing piece image is only 4 pixel sizes, is highly only 3 picturesElement size, as shown in Figure 5, each small circle represents a pixel, k representsColumns, m represents line number.

B. the quick slit cropping method of utilizing the yellow China of prior art to propose is set up between rowOptimum Matching relation, between row, the matching relationship of pixel only has 3 kinds of situations: a left sideUpper a, left side and lower-left, at matrix A R_3×4Element value corresponding to middle difference is-1,0,1.As shown in Figure 6, corresponding Optimum Matching relational matrix:

$\mathrm{AR}={\left[\begin{array}{cccc}0& 0& 0& 1\\ 0& 1& 1& -1\\ 0& -1& -1& 0\end{array}\right]}_{3\&times;4}$

C. utilize formula (6) to initialize finedraw label E (m):

E⁴(m)＝m,m＝1,2,3

E⁴(m) the finedraw label of capable the 4th row pixel of expression m, Fig. 7 is the 4th row pictureThe result figure of element mark. To neighborhood relationships matrix A, N initializes:

$\mathrm{AN}={\mathrm{AN}}^1={\mathrm{AN}}^2={\mathrm{AN}}^3={\left[\begin{array}{ccc}0& 0& 0\\ 0& 0& 0\\ 0& 0& 0\end{array}\right]}_{3\&times;3}$

D. utilize the finedraw label of formula (8) mark the 3rd row pixel, i.e. k=3:

E³(m+AR(m,3+1))＝E³⁺¹(m),m＝1,2,3

A) in the time of m=1, E³(1+AR(1,3+1))＝E³⁺¹(1)=1, wherein AR (Isosorbide-5-Nitrae)=1,E³(2)=1, is about to the 3rd pixel finedraw label parameter that is listed as the 2nd row and composesValue is 1;

B) in the time of m=2, E³(2+AR(2,3+1))＝E³⁺¹(2)=2, wherein AR (2,4)=-1,E³(1)=2, are about to the 3rd pixel finedraw label parameter that is listed as the 1st row and composeValue is 2;

C) in the time of m=3, E³(3+AR(3,3+1))＝E³⁺¹(3)=3, wherein AR (3,4)=0,E³(3)=3, are about to the 3rd pixel finedraw label parameter that is listed as the 3rd rowAssignment is 3;

Fig. 8 is the result figure of the 3rd row element marking, and Figure 10 is all pixel finedrawsLabel result figure.

E. utilize formula (9) constantly to upgrade neighborhood relationships matrix A N:

A) work as k=3, when m=1, p=(W-k) × h+m=(4-3) × 3+1=4

Bring formula (9) into:

AN⁴(E³(1),E³(2))＝AN³(E³(1),E³(2))+1

AN⁴(E³(2),E³(1))＝AN³(E³(2),E³(1))+1

Further:

AN⁴(2,1)＝AN³(2,1)+1＝0+1＝1

AN⁴(1,2)＝AN³(1,2)+1＝0+1＝1

Now,

$\mathrm{AN}={\mathrm{AN}}^4={\left[\begin{array}{ccc}0& 1& 0\\ 1& 0& 0\\ 0& 0& 0\end{array}\right]}_{3\&times;3}$

B) work as k=3, when m=2, p=(W-k) × h+m=(4-3) × 3+2=5

Bring formula (9) into:

AN⁵(E³(2),E³(1))＝AN⁴(E³(2),E³(1))+1

AN⁵(E³(1),E³(2))＝AN⁴(E³(1),E³(2))+1

AN⁵(E³(2),E³(3))＝AN⁴(E³(2),E³(3))+1

AN⁵(E³(3),E³(2))＝AN⁴(E³(3),E³(2))+1

Further:

AN⁵(1,2)＝AN⁴(1,2)+1＝1+1＝2

AN⁵(2,1)＝AN⁴(2,1)+1＝1+1＝2

AN⁵(1,3)＝AN⁴(1,3)+1＝0+1＝1

AN⁵(3,1)＝AN⁴(3,1)+1＝0+1＝1

Now,

$\mathrm{AN}={\mathrm{AN}}^5={\left[\begin{array}{ccc}0& 2& 1\\ 2& 0& 0\\ 1& 0& 0\end{array}\right]}_{3\&times;3}$

C) work as k=3, when m=3, p=(W-k) × h+m=(4-3) × 3+3=6

Bring formula (9) into:

AN⁶(E³(3),E³(2))＝AN⁵(E³(3),E³(2))+1

AN⁶(E³(2),E³(3))＝AN⁵(E³(2),E³(3))+1

Further:

AN⁶(3,1)＝AN⁵(3,1)+1＝1+1＝2

AN⁶(1,3)＝AN⁵(1,3)+1＝1+1＝2

Now,

$\mathrm{AN}={\mathrm{AN}}^6={\left[\begin{array}{ccc}0& 2& 2\\ 2& 0& 0\\ 2& 0& 0\end{array}\right]}_{3\&times;3}$

D) work as k=2, when m=1, p=(W-k) × h+m=(4-2) × 3+1=7

Bring formula (9) into:

AN⁷(E²(1),E²(2))＝AN⁶(E²(1),E²(2))+1

AN⁷(E²(2),E²(1))＝AN⁶(E²(2),E²(1))+1

Further:

AN⁷(2,3)＝AN⁶(2,3)+1＝0+1＝1

AN⁷(3,2)＝AN⁶(3,2)+1＝0+1＝1

Now,

$\mathrm{AN}={\mathrm{AN}}^7={\left[\begin{array}{ccc}0& 2& 2\\ 2& 0& 1\\ 2& 1& 0\end{array}\right]}_{3\&times;3}$

E) work as k=2, when m=2, p=(W-k) × h+m=(4-2) × 3+2=8

Bring formula (9) into:

AN⁸(E²(2),E²(1))＝AN⁷(E²(2),E²(1))+1

AN⁸(E²(1),E²(2))＝AN⁷(E²(1),E²(2))+1

AN⁸(E²(2),E²(3))＝AN⁷(E²(2),E²(3))+1

AN⁸(E²(3),E²(2))＝AN⁷(E²(3),E²(2))+1

Further:

AN⁸(3,2)＝AN⁷(3,2)+1＝1+1＝2

AN⁸(2,3)＝AN⁷(2,3)+1＝1+1＝2

AN⁸(3,1)＝AN⁷(3,1)+1＝2+1＝3

AN⁸(1,3)＝AN⁷(1,3)+1＝2+1＝3

Now,

$\mathrm{AN}={\mathrm{AN}}^8={\left[\begin{array}{ccc}0& 2& 3\\ 2& 0& 2\\ 3& 2& 0\end{array}\right]}_{3\&times;3}$

F) work as k=2, when m=3, p=(W-k) × h+m=(4-2) × 3+3=9

Bring formula (9) into:

AN⁹(E²(3),E²(2))＝AN⁸(E²(3),E²(2))+1

AN⁹(E²(2),E²(3))＝AN⁸(E²(2),E²(3))+1

Further:

AN⁹(1,3)＝AN⁸(1,3)+1＝3+1＝4

AN⁹(3,1)＝AN⁸(3,1)+1＝3+1＝4

Now,

$\mathrm{AN}={\mathrm{AN}}^9={\left[\begin{array}{ccc}0& 2& 4\\ 2& 0& 2\\ 4& 2& 0\end{array}\right]}_{3\&times;3}$

G) work as k=1, when m=1, p=(W-k) × h+m=(4-1) × 3+1=10

Bring formula (9) into:

AN¹⁰(E¹(1),E¹(2))＝AN⁹(E¹(1),E¹(2))+1

AN¹⁰(E¹(2),E¹(1))＝AN⁹(E¹(2),E¹(1))+1

Further:

AN¹⁰(2,1)＝AN⁹(2,1)+1＝2+1＝3

AN¹⁰(1,2)＝AN⁹(1,2)+1＝2+1＝3

Now,

$\mathrm{AN}={\mathrm{AN}}^{10}={\left[\begin{array}{ccc}0& 3& 4\\ 3& 0& 2\\ 4& 2& 0\end{array}\right]}_{3\&times;3}$

H) work as k=1, when m=2, p=(W-k) × h+m=(4-1) × 3+2=11

Bring formula (9) into:

AN¹¹(E¹(2),E¹(1))＝AN¹⁰(E¹(2),E¹(1))+1

AN¹¹(E¹(1),E¹(2))＝AN¹⁰(E¹(1),E¹(2))+1

AN¹¹(E¹(2),E¹(3))＝AN¹⁰(E¹(2),E¹(3))+1

AN¹¹(E¹(3),E1(2))＝AN¹⁰(E¹(3),E¹(2))+1

Further:

AN¹¹(1,2)＝AN¹⁰(1,2)+1＝3+1＝4

AN¹¹(2,1)＝AN¹⁰(2,1)+1＝3+1＝4

AN¹¹(1,3)＝AN¹⁰(1,3)+1＝4+1＝5

AN¹¹(3,1)＝AN¹⁰(3,1)+1＝4+1＝5

Now,

$\mathrm{AN}={\mathrm{AN}}^{11}={\left[\begin{array}{ccc}0& 4& 5\\ 4& 0& 2\\ 5& 2& 0\end{array}\right]}_{3\&times;3}$

I) work as k=1, when m=3, p=(W-k) × h+m=(4-1) × 3+3=12

Bring formula (9) into:

AN¹²(E¹(3),E¹(2))＝AN¹¹(E¹(3),E¹(2))+1

AN¹²(E¹(2),E¹(3))＝AN¹¹(E¹(2),E¹(3))+1

Further:

AN¹²(3,1)＝AN¹¹(3,1)+1＝5+1＝6

AN¹²(1,3)＝AN¹¹(1,3)+1＝5+1＝6

Now,

$\mathrm{AN}={\mathrm{AN}}^{12}={\left[\begin{array}{ccc}0& 4& 6\\ 4& 0& 2\\ 6& 2& 0\end{array}\right]}_{3\&times;3}$

AN is now final neighborhood relationships matrix.

3. calculate the neighbourship degree Neighborability between finedraw:

Neighbourship degree is used for representing adjacent probability between two finedraws, and it will be used for removing certainAfter bar finedraw, its adjacent finedraw is carried out to energy weighting. For example, to remove theM bar finedraw, we calculate its neighbourship degree sum and are:

$T\_\mathrm{AN}\left(m\right)=\underset{n=1}{\overset{h}{\mathrm{\&Sigma;}}}\mathrm{AN}(m,n)---\left(10\right)$

Then, calculate respectively other finedraws and the m article of probability that finedraw is adjacent. For example,

The probability calculation that n article of finedraw and m article of finedraw are adjacent is as follows:

$\mathrm{Neighborability}(m,n)=\frac{\mathrm{AN}(m,n)}{T\_\mathrm{AN}\left(m\right)},n=\mathrm{1,2},...h,m=\mathrm{1,2},...h---\left(11\right)$

4. cumlative energy and the neighbourship degree thereof of comprehensive finedraw are searched for finedraw to be processed:

In yellow method, only determine according to cumlative energy the finedraw that will remove. Accumulation energyMeasure littlely, more likely as removed finedraw, the present invention is tired by comprehensive finedrawLong-pending energy and neighbourship degree determine the finedraw that will remove, and concrete steps are as follows:

A. remove the finedraw m of cumlative energy minimum according to least energy principle.

B. then use formula (12) to upgrade its cumlative energy to remaining finedraw

$\stackrel{\&OverBar;}{\mathrm{AE}\left(n\right)}=\mathrm{AE}\left(n\right)+\mathrm{AE}\left(n\right)*w\left(n\right),n=\mathrm{1,2},...h---\left(12\right)$

In formula (12), AE (n) represents the cumlative energy of n article of finedraw, and w (n) is specifically fixed

Justice is:

w(n)＝C*Neighborability(m,n),n＝1,2...h(13)

In formula (13), C is a constant, and value is 1 in this example. Be used for adjustingThe influence degree that joint neighbourship degree distributes to finedraw. Neighborability (m, n) tableShow the neighbourship degree of n article of finedraw and m article of finedraw.

C. repeating step A and B, we can find the Num that all will remove_i=h_i’-hThe optimum finedraw of bar, removes corresponding finedraw, obtains object height strips S trips.

Claims (1)

1. the quick slit cropping method based on band and neighbourship degree, is characterized in that: shouldMethod specifically comprises:

1) input one width size is the original image of W*H, and sets its target size and beW*H_T, W is picture traverse, the height that H is original image, H_TFor the height of target image;

2) original image is divided into the equidistant horizontal strip of N bar;

3) calculate the importance S of each band according to Saliency maps corresponding to former figure_i；

4) according to the importance values S of each band_iAnd the height H of target image_TGo to calculateThe object height h of each band_i’；

5) according to the original height h of each band and object height h_i', can determine eachBand needs the finedraw quantity Num removing_i＝h_i'-h; Then in each band according to neighbourship degreeSearch for finedraw to be processed with cumulative energy;

Described step 5) be specially: obtain target size image by deleting these finedraws,The concrete steps of each band search finedraw are as follows:

1. the Optimum Matching relation between calculating pixel:

Adopt quick slit cropping method that the yellow China of prior art proposes by between adjacent columnsThe weight of edges matched and set up the Optimum Matching between pixel in original image between largeization pixelRelation; Between every two row, carry out iterative computation determine be listed as between the Optimum Matching relation of pixel, enterAnd remove to obtain the Optimum Matching relational matrix AR of whole band_W×h；

2. calculate the neighborhood relationships between finedraw according to pixel Optimum Matching relation:

The horizontal strip that is h for a height, we need to search for h bar finedraw; Definition is thinBe sewn on k and be listed as its finedraw label of pixel element E that m is capable^k(m) mark; In addition, weDefinition accumulation neighborhood relationships matrix A N_h×h, this matrix is weighing according to two finedraw elements of each rowWant whether region is adjacent accumulates renewal; Important area is averaged by the remarkable figure of imageValue is carried out binaryzation as threshold value again and is obtained; The process of setting up finedraw neighborhood relationships matrix A N fromRight-to-left is carried out, and on the right side, one row are k=W, and the finedraw label parameter of pixel is initialized as:

E^W(m)＝m,m＝1,2,...h(3)

Neighborhood relationships matrix A N is initialized as:

AN(x,y)＝0,x＝1,2,...hy＝1,2,...h(4)

Further be related to that according to the Optimum Matching between pixel AR is according to following Policy Updates finedraw labelParameters E (m) and neighborhood relationships matrix A N:

E^k(m+AR(m,k+1))＝E^k+1(m),m＝1,2,...h,k＝1,2,...W(5)

$\left\{\begin{array}{c}{\mathrm{AN}}^p\left(E^k\right(m),E^k(m-1))={\mathrm{AN}}^{p-1}\left(E^k\right(m),E^k(m-1))+1,k=W-1,...1,m=2,...h\\ {\mathrm{AN}}^p\left(E^k\right(m-1),E^k\left(m\right))={\mathrm{AN}}^{p-1}\left(E^k\right(m-1),E^k\left(m\right))+1,k=W-1,...1,m=2,...h\\ {\mathrm{AN}}^p\left(E^k\right(m),E^k(m+1))={\mathrm{AN}}^{p-1}\left(E^k\right(m),E^k(m+1))+1,k=W-1,...1,m=1,...h-1\\ {\mathrm{AN}}^p\left(E^k\right(m+1),E^k\left(m\right))={\mathrm{AN}}^{p-1}\left(E^k\right(m+1),E^k\left(m\right))+1,k=W-1,...1,m=1,...h-1\\ p=(W-k)\&times;h+m\end{array}\right.---\left(6\right)$

In formula (5), AR (m, k) is the capable k column position of m place in Optimum Matching relational matrix ARValue, E^k(m) be the finedraw number value of the capable k row of m pixel; In formula (6), AN^pRepresent to have upgradedThe value of neighborhood relationships matrix A N when p pixel, p=1 represents band right column the first rowPixel; Until be updated to last pixel element of first row, we can obtain final neighbourTerritory relational matrix AN_h×h；

3. calculate the neighbourship degree Neighborability between finedraw:

Neighbourship degree is used for representing adjacent probability between two finedraws, and it will be used for removing certain carefullyAfter seam, its adjacent finedraw is carried out to energy weighting; To the m article of finedraw of removing, Wo MenjiCalculating its neighbourship degree sum is:

$T\_AN\left(m\right)=\underset{n=1}{\overset{h}{\&Sigma;}}AN(m,n)---\left(7\right)$

Then, calculate respectively other finedraws and the m article of probability that finedraw is adjacent; N article of finedrawThe probability calculation adjacent with m article of finedraw is as follows:

$Neighborability(m,n)=\frac{AN(m,n)}{T\_AN\left(m\right)},n=1,2,...h,m=1,2,...h---\left(8\right)$

4. cumlative energy and the neighbourship degree thereof of comprehensive finedraw are searched for finedraw to be processed:

A. remove the finedraw m of cumlative energy minimum according to least energy principle;

B. then use formula (9) to upgrade its cumlative energy to remaining finedraw,

In formula (9), AE (n) represents the cumlative energy of n article of finedraw, and w (n) is according to neighbourship degree meterGet, specific definition is:

w(n)＝C*Neighborability(m,n),n＝1,2...h(10)

In formula (10), C is a constant, and value is 1; Be used for regulating neighbourship degree to distribute to finedrawInfluence degree; Neighborability (m, n) represents n article of finedraw and m article of finedrawNeighbourship degree;

C. repeating step A and B, finds the Num that all will remove_i＝h-h_iThe optimum finedraw of bar, goesExcept corresponding finedraw, obtain the horizontal strip of object height.