CN103778598B - Disparity map ameliorative way and device - Google Patents

Abstract

Providing a kind of disparity map ameliorative way and device, the method may include that acquisition initial parallax figure；Determine the similarity between the pixel of disparity map；Building graph model, wherein each node of figure corresponds to each image pixel, and the initial value of each node is the initial parallax value of respective pixel, and in figure, the limit between node has the corresponding weight of the similarity between respective pixel；The parallax value of node in figure is diffused iteration, in order to each wheel in diffusion iteration, the parallax value of each node being broadcast to other node, communication process terminates at convergence or reaches a certain maximum iteration time；And using the parallax value of node in the figure after propagation termination as the parallax value of respective pixel.Utilize this disparity map ameliorative way and device, all pixels in image are built graph model, reach global optimum by the successive ignition propagation in graph model, noise and accentuated edges can be reduced；And without image is split, calculating resource can be saved.

Description

Disparity map ameliorative way and device

Technical field

The present invention relates generally to image procossing, relates more specifically to disparity map ameliorative way and device.

Background technology

Images match is one of the basic research problems in computer vision and image understanding field, is widely used to the numerous areas such as multi-source image fusion, target recognition, three-dimensional reconstruction.

Along with the rise in the field such as three-dimensional reconstruction, virtual reality, people are more and more higher to the precision of disparity map and the requirement of density, it is desirable to the reliable depth information of each pixel in acquisition image.The purpose of Stereo matching is exactly to find the pixel pair of coupling in two width imaging planes of identical three-dimensional scenic, calculates the parallax value that each pixel coordinate is corresponding, and then can describe the three-dimensional spatial information of scene accurately.

Patent documentation No.US20050286756A1 discloses a kind of parallax calculation method/system based on image segmentation.First image is split by the method, and follow-up improvement is to use various matching cost equations, and estimates disparity plane by continuous iteration, thus obtains more accurate parallax value.In the method, in order to obtain more accurate parallax value, surface fitting repeatedly is necessary, thus can consume the more calculating time.On the other hand, this parallax ameliorative way is not an overall method, so usually can only obtain preferable effect in some region of image.

HeikoHirschmuller, at entitled " StereoprocessingbySemi-GlobalMatchingandMutualInformatio n, " inTPAMI, proposes the disparity correspondence method of one and half overall situations in the non-patent literature of 2008.This article utilize peak filtering (Peaksremoval) and plane fitting (Planefitting) method carry out the initial parallax value obtained improving optimization.This method is not global approach, is difficult to obtain preferable effect in full image, and additionally the method needs to split image in advance, and this will consume the substantial amounts of time.

Summary of the invention

According to embodiments of the invention, it is provided that a kind of disparity map ameliorative way, may include that acquisition initial parallax figure；Determine the similarity between the pixel of disparity map；Building graph model, wherein each node of figure corresponds to each image pixel, and the initial value of each node is the initial parallax value of respective pixel, and in figure, the limit between node has the corresponding weight of the similarity between respective pixel；The parallax value of node in figure is diffused iteration, in order to each wheel in diffusion iteration, the parallax value of each node being broadcast to other node, communication process terminates at convergence or reaches a certain maximum iteration time；And using the parallax value of node in the figure after propagation termination as the parallax value of respective pixel.

According to another embodiment of the present invention, it is provided that a kind of disparity map improves device, may include that initial parallax figure obtains parts, it is thus achieved that initial parallax figure；Between pixel, similarity determines parts, determines the similarity between the pixel of disparity map；Graph model structural member, builds graph model, and wherein each node of figure corresponds to each image pixel, and the initial value of each node is the initial parallax value of respective pixel, and in figure, the limit between node has the corresponding weight of the similarity between respective pixel；Diffusion iteration parts, are diffused iteration to the parallax value of node in figure, in order to each wheel in diffusion iteration, the parallax value of each node being broadcast to other node, communication process terminates at convergence or reaches a certain maximum iteration time；And final parallax obtains parts, in the figure after terminating using propagation, the parallax value of node is as the parallax value of respective pixel.

Utilize disparity map ameliorative way according to embodiments of the present invention and device, all pixels in image are built graph model, reach global optimum by the successive ignition propagation in graph model, noise and accentuated edges can be reduced；And without image is split, calculating resource can be saved.

Accompanying drawing explanation

Fig. 1 schematically shows parallax according to an embodiment of the invention and improves exemplary configuration and the operation of system.

Fig. 2 shows the overview flow chart of parallax ameliorative way according to a first embodiment of the present invention.

Fig. 3 schematically shows the gray scale difference value between pixel and surrounding pixel and the relation between similarity calculated based on formula (1).

Fig. 4 shows that disparity map according to embodiments of the present invention improves the functional configuration block diagram of device.

Fig. 5 is the general hardware block diagram illustrating and improving system according to the disparity map of the embodiment of the present invention.

Detailed description of the invention

In order to make those skilled in the art be more fully understood that the present invention, with detailed description of the invention, the present invention is described in further detail below in conjunction with the accompanying drawings.

To be described in the following order:

1, parallax improves configuration and the operational overview of system

2, disparity map ameliorative way

3, disparity map improves device

4, system hardware configuration

5, sum up

1, parallax improves configuration and the operational overview of system

It is broadly described the parallax mated towards stereoscopic vision according to an embodiment of the invention below with reference to Fig. 1 and improves exemplary configuration and the operation of system.

Fig. 1 schematically shows parallax according to an embodiment of the invention and improves exemplary configuration and the operation of system.

In the example depicted in fig. 1, parallax improves system 100 and includes: binocular camera 101, obtain left Figure 102 (1) and right Figure 102 (2) by binocular camera 101 shooting, or 102 can represent reference picture 102 (1) that is that binocular photographing unit exports and that pass through correction process and coupling image 102 (2)；Initial parallax figure obtains parts (not shown), is calculated initial parallax Figure 103 based on reference picture and coupling image；Graph model builds parts 104, for pixel is built graph model, wherein each node of figure corresponds to each image pixel, and the initial value of each node is the initial parallax value of respective pixel, and in figure, the limit between node has the corresponding weight of the similarity between respective pixel；Parallax value iterative diffusion parts 105, the parallax value of node in figure is diffused iteration, the parallax value of each node to be broadcast to other node (certainly each wheel in diffusion iteration, the parallax value of each node is adjusted also by the propagation of the parallax value of other node), communication process terminates at convergence or reaches a certain maximum iteration time, and communication process will obtain final parallax Figure 106 when terminating.Graph model builds parts 104 and parallax value iterative diffusion parts 105 figure 3 illustrates as being realized by software program or the firmware installed on chip 107, but is only for example, those skilled in the art will know that it can also be realized by specialised hardware.

2, disparity map ameliorative way

Fig. 2 shows the overview flow chart of parallax ameliorative way 200 according to a first embodiment of the present invention.

As in figure 2 it is shown, in step S210, it is thus achieved that initial parallax figure.Initial parallax figure can obtain based on the coupling between reference picture and coupling image, and the method for various traditional calculating parallax value can be utilized to draw, such as Block-matching, dynamic programming, figure cut method, half global registration method (SGM) etc..

In step S220, determine the similarity between the pixel of disparity map.

Exemplarily, can one or more similarities determined between pixel in the feature such as gray value based on the pixel in reference picture, color value, texture value.The most by the way, reference picture and coupling image role be interchangeable, i.e. coupling image can as reference picture, and reference picture be used as coupling image.

It addition, during similarity between the pixel determining disparity map, it may be considered that the distance between pixel.Such as, just consider the calculating of similarity at a distance of the pixel within preset range, and be simply considered that not there is similarity between the pixel exceeding preset range apart.

In follow-up graph model in the diffusion iteration of the parallax value of node, do not have between the pixel of similarity and will not influence each other.

For example, it is possible to the similarity between the pixel of disparity map is configured to similarity matrix W (n × n), wherein n is the number of pixels of reference picture, value w of the i-th row jth row of similarity matrix_ijRepresent the similarity between pixel i and pixel j in disparity map.For example, it is possible to calculate similarity w between pixel i and pixel j according to following formula (1)_ij,

D (i, j) represents the difference between pixel i and the characteristic vector of pixel j, such as, squared differences and, absolute difference and etc., the element of characteristic vector can comprise gray scale, color, texture etc..In one example, d (i, j) represents the absolute difference between certain characteristic element, i.e. d (i, j)=| I_i-I_j|, I_iAnd I_jReferring respectively to pixel i and one of the gray value of j, color value, texture value, σ is scale parameter, for regulating the scope of data distribution.

It should be noted that in this article, two pixel is neighbouring or pixel i is identical in the neighborhood implication of pixel j, such as between two pixels, the distance in X-Y coordinate is less than predetermined threshold.

It addition, above-mentioned formula (1) have employed Gauss model form to the Similarity Measure of neighborhood pixels, this is in order to pixel separability is higher.It addition, in formula (1), σ is the radius parameter in Gauss model, for adjusting the degree of separability.The value of σ can be by experimental verification or utilize the means of statistics to obtain, as an example, and can be it is set to the intermediate value of gray scale difference value of all pixels.

The similarity between pixel is evaluated based on formula (1), and using this similarity as the weight on limit in graph model, and then carry out the parallax value of node in figure to be iterated propagation being based on considering as follows: for two neighbouring pixels, if the feature of the gray level images such as gray scale, color or texture is closer to, then to be likely to comparison similar for the parallax value of these two pixels, therefore should interact.

Fig. 3 schematically shows the gray scale difference value between pixel and surrounding pixel and the relation between similarity calculated based on formula (1).

(a) in Fig. 3 represents the coloured image (because not allowing color occur in patent document, therefore behave as gray-scale map) of reference picture, and (b) represents the zoomed-in view of the part in the dotted line frame in coloured image.Pixel p is a pixel randomly choosed, and pixel a and b are two neighbouring pixels of pixel p.According to the definition of formula (1), (p, a) more than the gray scale difference d (p between pixel p and pixel b for the gray scale difference d between pixel p and pixel a, b), i.e. d (p, a)>d (p, b), thus can obtain: the similarity w (p between pixel p and pixel a, a) less than between pixel p and pixel b similarity d (p, b), i.e. w (p, a)<w (p, b).

Note that in figure 3, with regard to pixel x, for the distance on y-coordinate, the distance between pixel p and pixel a is equal to the distance between pixel p and pixel b.According to some prior art, may only consider the distance between pixel coordinate to determine which pixel parallax influences each other and effect, so can draw w (p, a)=w (and p, b).

Comparatively speaking, the method for the embodiment of the present invention more tallies with the actual situation.Because it practice, parallax (degree of depth) difference of the parallax of pixel p (in other words, the degree of depth) and pixel a is relatively big, and the parallax (degree of depth) of the parallax of pixel p (in other words, the degree of depth) and pixel a is the most close.Therefore, for parallax improves, only consider that the distance between pixel coordinate determines which pixel parallax influences each other compared to former, the embodiment of the present invention consider that similarity that the gray scale of gray level image, color or Texture eigenvalue investigate between pixel is to determine which pixel influences each other parallax more reasonability each other.

It should be noted that above-mentioned formula (1) is merely illustrative.Those skilled in the art can select the different characteristic of pixel as required and/or utilize different formula form to the difference calculating between pixel, and and then calculate the similarity between pixel.More generally, using the coordinate figure of pixel, gray scale, color value, texture value etc. as the characteristic element of pixel, the characteristic vector of composition pixel, by the difference between the characteristic vector of calculating pixel (such as, Euclidean distance, absolute difference and etc.) weigh the similarity between pixel, additionally when calculating the similarity between pixel, each characteristic element can have different weight factors.

Preferably, similarity matrix W can be normalized so that in the range of the value of all elements in similarity matrix W is respectively positioned on [0,1].

For example, it is possible to similarity matrix W is normalized by following formula (2).

W=N^-1/2WN^-1/2...(2)

Here matrix N is a diagonal matrix, its i-th row i-th column element the i-th row element sum equal to similarity matrix W.

As another example, following formula (3) can be passed through and be normalized.

W=N^-1W…(3)

In step S230, building graph model, wherein each node of figure corresponds to each image pixel, and the initial value of each node is the initial parallax value of respective pixel, and in figure, the limit between node has the corresponding weight of the similarity between respective pixel.

In step S240, the parallax value of node in figure is diffused iteration, in order to each wheel in diffusion iteration, the parallax value of each node being broadcast to other node, communication process terminates at convergence or reaches a certain maximum iteration time.

D(t+1)=αWD(t)+(1-α)D(0)…(4)

The parallax value matrix of n × 1 when wherein D (t) is the t time iteration, in matrix, the value of each element is corresponding to the parallax value of each pixel in n pixel；D (0) is initial parallax value matrix；Real number α is weighting parameter, and for controlling the contribution amount of initial parallax, 0 < α < 1, α is the least, represents that initial parallax value is the most credible；W is similarity matrix or weight matrix.

In step s 250, in the figure after terminating using propagation, the parallax value of node is as the parallax value of respective pixel.

Hereinafter prove that iterative process is convergence.

In the case of not affecting result, simple for representing, if D (0)=Y, following formula (5) can be derived according to formula (4)

$D\left(t\right)={\left(\mathrm{\&alpha;W}\right)}^{t-1}Y+(1-\mathrm{\&alpha;}){\mathrm{\&Sigma;}}_{i=0}^{t-1}(\mathrm{\&alpha;W}{)}^{i}Y...\left(5\right)$

In view of 0 < α < 1, and the characteristic value of matrix W is in the range of scope [-1,1], can obtain formula (6)

lim_t→∞(αW)^t-1=0…(6)

Can obtain further,

${\lim }_{t\&RightArrow;\&infin;}{\mathrm{\&Sigma;}}_{i=0}^{t-1}{\left(\mathrm{\&alpha;W}\right)}^i={(I-\mathrm{\&alpha;S})}^{-1}...\left(7\right)$

Therefore, if parallax matrix finally converges on D^*, D* can be represented by following formula (8)

$D^{*}=\underset{t\&RightArrow;\&infin;}{\lim }D\left(t\right)=(1-\mathrm{\&alpha;}){(I-\mathrm{\&alpha;S})}^{-1}Y...\left(8\right)$

Visible final parallax value has unique solution.

Utilize disparity map ameliorative way according to embodiments of the present invention, all pixels in image are built graph model, reach global optimum by the successive ignition propagation in graph model, noise and accentuated edges can be reduced；And without image is split, calculating resource can be saved.

3, disparity map improves device

The disparity map described according to embodiments of the present invention below with reference to Fig. 4 improves device.

Fig. 4 shows that disparity map according to embodiments of the present invention improves the functional configuration block diagram of device 400.

As shown in Figure 4, match measure determines that device 400 may include that initial parallax figure obtains parts 410, it is thus achieved that initial parallax figure；Between pixel, similarity determines parts 420, determines the similarity between the pixel of disparity map；Graph model structural member 430, builds graph model, and wherein each node of figure corresponds to each image pixel, and the initial value of each node is the initial parallax value of respective pixel, and in figure, the limit between node has the corresponding weight of the similarity between respective pixel；Diffusion iteration parts 440, are diffused iteration to the parallax value of node in figure, in order to each wheel in diffusion iteration, the parallax value of each node being broadcast to other node, communication process terminates at convergence or reaches a certain maximum iteration time；And final parallax obtains parts 450, in the figure after terminating using propagation, the parallax value of node is as the parallax value of respective pixel.

Above-mentioned initial parallax figure obtains similarity between parts 410, pixel and determines that parts 420, graph model structural member 430, diffusion iteration parts 440, the concrete function of final parallax acquisition parts 450 are referred to the above-mentioned description relevant with Fig. 1 to Fig. 3 with operation.Here relevant repeated description is omitted.

4, system hardware configuration

The present invention can also improve hardware system by a kind of disparity map and implement.Fig. 5 is the general hardware block diagram illustrating and improving system 1000 according to the disparity map of the embodiment of the present invention.As shown in Figure 5, disparity map improves system 1000 and may include that input equipment 1100, for inputting relevant image or information from outside, the left image of such as video camera shooting and right image, the parameter of video camera or initial parallax figure etc., such as, can include keyboard, Genius mouse and communication network and remote input equipment of being connected thereof etc.；Processing equipment 1200, method is determined for implementing the above-mentioned match measure according to the embodiment of the present invention, or it is embodied as above-mentioned match measure and determines device, or implement above-mentioned disparity map ameliorative way, central processing unit or other the chip with disposal ability that such as can include computer etc., may be coupled to the network (not shown) of such as the Internet, need the image etc. after teletransmission processes according to processing procedure；Outut device 1300, improves the result of process gained for implementing above-mentioned disparity map to outside output, such as, can include display, printer and communication network and remote output devices of being connected thereof etc.；And storage device 1400, the data such as the disparity map after storing such as initial parallax figure involved by above-mentioned disparity map improvement process, similarity matrix, graph model, improvement in the way of volatile and nonvolatile, such as, can include the various volatile and nonvolatile property memorizer of random-access memory (ram), read only memory (ROM), hard disk or semiconductor memory etc..

5, sum up

According to embodiments of the invention, it is provided that a kind of disparity map ameliorative way, may include that acquisition initial parallax figure；Determine the similarity between the pixel of disparity map；Building graph model, wherein each node of figure corresponds to each image pixel, and the initial value of each node is the initial parallax value of respective pixel, and in figure, the limit between node has the corresponding weight of the similarity between respective pixel；The parallax value of node in figure is diffused iteration, in order to each wheel in diffusion iteration, the parallax value of each node being broadcast to other node, communication process terminates at convergence or reaches a certain maximum iteration time；And using the parallax value of node in the figure after propagation termination as the parallax value of respective pixel.

According to another embodiment of the present invention, it is provided that a kind of disparity map improves device, may include that initial parallax figure obtains parts, it is thus achieved that initial parallax figure；Between pixel, similarity determines parts, determines the similarity between the pixel of disparity map；Graph model structural member, builds graph model, and wherein each node of figure corresponds to each image pixel, and the initial value of each node is the initial parallax value of respective pixel, and in figure, the limit between node has the corresponding weight of the similarity between respective pixel；Diffusion iteration parts, are diffused iteration to the parallax value of node in figure, in order to each wheel in diffusion iteration, the parallax value of each node being broadcast to other node, communication process terminates at convergence or reaches a certain maximum iteration time；And final parallax obtains parts, in the figure after terminating using propagation, the parallax value of node is as the parallax value of respective pixel.

Utilize utilization according to embodiments of the present invention disparity map ameliorative way according to embodiments of the present invention and device, all pixels in image are built graph model, propagated by the successive ignition in graph model and reach global optimum, noise and accentuated edges can be reduced；And without image is split, calculating resource can be saved.

Described above the most illustrative, can much revise and/or replace.

Accompanying drawing above and description mate with binocular stereo vision, i.e. illustrates as a example by left images coupling, but the invention is not limited in this, but any images match for stereoscopic vision purpose can be applied.

Additionally, described above it is calculated initial parallax figure based on left images, the most how to obtain initial parallax figure non-invention emphasis, the initial parallax figure that any method or approach obtain may be incorporated for the present invention, and such as initial parallax figure can be from remotely obtaining.

Additionally, in example above, when the similarity calculated between pixel, for to each other in x-y coordinate system distance exceed predetermined threshold, similarity is simply set as zero, but is only for example, as previously described, this distance factor can be otherwise account for so that distance is the nearest, and similarity is the biggest in Similarity Measure.

It should be noted that parallax or disparity map should be from broadly understanding in the present invention, it also comprises the degree of depth or depth map, therefore it will be apparent to those skilled in the art that and can be by between parallax and the degree of depth being simply mutually converted to.

The ultimate principle of the present invention is described above in association with specific embodiment, but, it is to be noted, for those of ordinary skill in the art, it will be appreciated that whole or any steps of methods and apparatus of the present invention or parts, can be in any calculating device (including processor, storage medium etc.) or the network calculating device, being realized with hardware, firmware, software or combinations thereof, this is that those of ordinary skill in the art use their basic programming skill can be achieved with in the case of the explanation having read the present invention.

Therefore, the purpose of the present invention can also be realized by one program of operation or batch processing on any calculating device.Described calculating device can be known fexible unit.Therefore, the purpose of the present invention can also realize only by providing the program product comprising the program code realizing described method or device.It is to say, such program product also constitutes the present invention, and storage has the storage medium of such program product also to constitute the present invention.Obviously, described storage medium can be any known storage medium or any storage medium developed in the future.

It may also be noted that in apparatus and method of the present invention, it is clear that each parts or each step can decompose and/or reconfigure.These decompose and/or reconfigure the equivalents that should be regarded as the present invention.Further, the step performing above-mentioned series of processes can order the most following the instructions perform in chronological order, but is not required to perform the most sequentially in time.Some step can perform parallel or independently of one another.

Above-mentioned detailed description of the invention, is not intended that limiting the scope of the invention.Those skilled in the art, it is to be understood that depend on that design requires and other factors, can occur various amendment, combination, sub-portfolio and replacement.Any amendment, equivalent and improvement etc. made within the spirit and principles in the present invention, within should be included in scope.

Claims (8)

1. a disparity map ameliorative way, including:

Obtain initial parallax figure；

Determine the similarity between the pixel of disparity map；

Building graph model, wherein each node of figure corresponds to each image pixel, and the initial value of each node is the initial parallax value of respective pixel, and in figure, the limit between node has the corresponding weight of the similarity between respective pixel；

The parallax value of node in figure is diffused iteration, in order to each wheel in diffusion iteration, the parallax value of each node being broadcast to other node, communication process terminates at convergence or reaches a certain maximum iteration time；And

In figure after terminating using propagation, the parallax value of node is as the parallax value of respective pixel；

Wherein by following propagation formula, the parallax value of node in figure is diffused iteration,

D (t+1)=α WD (t)+(1-α) D (0)

The parallax value matrix of n × 1 when wherein D (t) is the t time iteration, in matrix, the value of each element is corresponding to the parallax value of each pixel in n pixel, and D (0) is initial parallax value matrix, and real number α is weighting parameter, for controlling the contribution amount of initial parallax, 0 < α < 1.

Disparity map ameliorative way the most according to claim 1, wherein initial parallax figure is to obtain based on the coupling between reference picture and coupling image, and one or more in the gray value based on the pixel in reference picture of the similarity between the pixel of disparity map, color value, texture value determine.

Disparity map ameliorative way the most according to claim 2, wherein the similarity between the pixel of disparity map is configured to similarity matrix W (n × n), and wherein n is the number of pixels of reference picture, value w of the i-th row jth row of similarity matrix_ijRepresent the similarity between pixel i and pixel j, similarity w between pixel i calculated as below and pixel j in disparity map_ij,

I_iAnd I_jReferring respectively to pixel i and one of the gray value of j, color value, texture value, σ is scale parameter, for regulating the scope of data distribution,

Wherein in figure the weight on limit between the i-th node and jth node equal to similarity w between pixel i and pixel j_ij。

Disparity map ameliorative way the most according to claim 3, also includes being normalized similarity matrix W.

5. disparity map improves a device, including:

Initial parallax figure obtains parts, it is thus achieved that initial parallax figure；

Between pixel, similarity determines parts, determines the similarity between the pixel of disparity map；

Graph model structural member, builds graph model, and wherein each node of figure corresponds to each image pixel, and the initial value of each node is the initial parallax value of respective pixel, and in figure, the limit between node has the corresponding weight of the similarity between respective pixel；

Diffusion iteration parts, are diffused iteration to the parallax value of node in figure, in order to each wheel in diffusion iteration, the parallax value of each node being broadcast to other node, communication process terminates at convergence or reaches a certain maximum iteration time；And

Final parallax obtains parts, and in the figure after terminating using propagation, the parallax value of node is as the parallax value of respective pixel；

Wherein by following propagation formula, the parallax value of node in figure is diffused iteration,

D (t+1)=α WD (t)+(1-α) D (0)

The parallax value matrix of n × 1 when wherein D (t) is the t time iteration, in matrix, the value of each element is corresponding to the parallax value of each pixel in n pixel, and D (0) is initial parallax value matrix, and real number α is weighting parameter, for controlling the contribution amount of initial parallax, 0 < α < 1.

Disparity map the most according to claim 5 improves device, wherein initial parallax figure is to obtain based on the coupling between reference picture and coupling image, and one or more in the gray value based on the pixel in reference picture of the similarity between the pixel of disparity map, color value, texture value determine.

Disparity map the most according to claim 6 improves device, and wherein the similarity between the pixel of disparity map is configured to similarity matrix W (n × n), and wherein n is the number of pixels of reference picture, value w of the i-th row jth row of similarity matrix_ijRepresent the similarity between pixel i and pixel j, similarity w between pixel i calculated as below and pixel j in disparity map_ij,

I_iAnd I_jReferring respectively to pixel i and one of the gray value of j, color value, texture value, σ is scale parameter, for regulating the scope of data distribution,

Wherein in figure the weight on limit between the i-th node and jth node equal to similarity w between pixel i and pixel j_ij。

Disparity map the most according to claim 7 improves device, also includes being normalized similarity matrix W.