CN105007478A - View synthesis method based on mean shift stereo matching - Google Patents

Abstract

The present invention discloses a view synthesis method based on mean shift stereo matching. The method comprises: firstly, considering two commonly used images including gray images and color images, a mean shift image segmentation method is used for cutting segments in a known reference image; secondly, based on a weight multi-window matching method based on color similarity between pixels and an introduced matching cost function, an initial disparity value is optimized to gain a more accurate disparity map of a stereo image pair, and the disparity map is smoothed; and at last, positive view interpolation and hollow noise processing is conducted, and finally image drawing at any position between a left view and a right view. The method gains good stereo effects, and is suitable for gray images and color images.

Description

A kind of view synthesizing method based on mean shift Stereo matching

Technical field

The present invention relates to a kind of view synthesizing method, particularly relate to a kind of view synthesizing method based on mean shift Stereo matching, belong to picture processing field.

Background technology

Along with the fast development in digital broadcast television technology and three-dimensional movie market, traditional two dimensional image can not meet the personal requirement of people to visually-perceptible far away, so three-dimensional television (3DTV) has become " best pet " that people pursue visual stereoscopic sense and the sense of reality.3DTV and FTV is various visual angles view (MVI) most important applications, and in 3DTV and FTV, virtual view synthetic technology is particularly important core technology.Nowadays, drawing virtual view image technology has become a study hotspot of digital picture and computer vision field.

In prior art, some scholars propose several method for drawing virtual view image, but all there is certain defect: the geometry in certain window of giving chapter and verse in Locally adaptive support-weight approach for visualcorrespondence search. (the IEEE Transactions on Pattern Analysis and MachineIntelligence.2006) literary composition that Yoon KJ and Kweon I S. delivers between pixel and point to be matched and photometric relationship are to adjust the weights of each pixel, the method has good anti-interference and robustness to the matching result becoming window, but it is too large that the shortcoming of the method is amount of calculation, and weight function used is not suitable for the coupling of gray-scale map, lack degree of widely using.Kanade T and Okutomi M. proposes the method by changing window size and shape, make each pixel can obtain best match window value, namely the window value calculated when Matching power flow function makes its value reach minimum is the best, and the method is for the pixel poor effect in the discontinuous region of the degree of depth.

Summary of the invention

Technical problem to be solved by this invention provides a kind of view synthesizing method based on mean shift Stereo matching for the deficiency of background technology.

The present invention is for solving the problems of the technologies described above by the following technical solutions

Based on a view synthesizing method for mean shift Stereo matching, specifically comprise the steps:

Step 1, adopts mean shift method to carry out Region Segmentation to original image;

Step 2, adopts the multiwindow stereo matching method based on adaptive weight to obtain disparity map corresponding to stereo-picture;

Step 3, to the smoothing operation of step 2 gained disparity map;

Step 4, carries out forward view interpolation and empty noise processed to disparity map after smooth operation, and then completes Image Rendering.

As the further preferred version of a kind of view synthesizing method based on mean shift Stereo matching of the present invention, described step 1 detailed process is as follows:

Step 1.1, is converted to the chrominance space matched with image by image;

Step 1.2, utilizes the smoothing operation of mean shift and then obtains the convergence point of each pixel;

Step 1.3, merges the convergence point of each pixel according to the conditional plan preset;

As the further preferred version of a kind of view synthesizing method based on mean shift Stereo matching of the present invention, described step 2 detailed process is as follows:

Step 2.1, certain pixel constructs one and supports window in a reference image;

Step 2.2, moves in parallel this support window, carrys out the similitude between calculation window with similarity measure function along disparity range; Specifically be calculated as follows:

$C_{SAD}(x,y)=\underset{c\∈\{r,g,b\}}{\Σ}|I_c\left(x\right)-I_c\left(y\right)|$

$\begin{array}{c}{C}_{GRAD}(x,y)=\underset{c\∈\left\{r,g,b\right\}}{\Σ}{\left|{\▿}_x{I}_c\left(x\right)-{\▿}_x{I}_c\left(y\right)\right|}^2+\\ \underset{c\∈\left\{r,g,b\right\}}{\Σ}{\left|{\▿}_y{I}_c\left(x\right)-{\▿}_y{I}_c\left(y\right)\right|}^2\end{array}$

C(x,y)＝(1-w)*C _GRAD(x,y)+w*C _SAD(x,y)

Wherein, C _sAD(x, y) is for asking for the absolute value sum of current pixel point and field pixel r, g, b triple channel color distortion, I _cx () is current pixel point value of color, x is current point coordinate, I _c(y) for field value of color around current pixel point, y be world coordinates, with the horizontal and vertical gradient of representative image respectively, w is the weights between 0 and 1, C _gRAD(x, y) for ask for current point and its field pixel in the horizontal direction with the absolute value sum of vertical direction r, g, b triple channel color distortion, C (x, y) is similarity measure function.

As the further preferred version of a kind of view synthesizing method based on meanshift Stereo matching of the present invention, described step 4 detailed process is as follows:

Step 4.1, obtains positional information and the colouring information of each pixel;

Step 4.2, then by the disparity map determination virtual view position relationship of this view, determine evolution relation, with correspondence former depending on colouring information color filling is carried out to virtual view, forward view interpolation formula is as follows:

I _IR(X _I,Y)＝I _IR(X _R+(1-α)*d ^RL,Y)＝I _R(X _R,Y)

Wherein, X _r+ (1-α) * d ^rLfor virtual view position and right view position relationship, I _iR(X _i, Y) and for transform to intermediate virtual viewpoint pixel coordinate by right view be (X _i, Y) and the pixel value at place, d ^rLfor the parallax value of left and right view, I _r(X _r, Y) for right view pixel coordinate be (X _r, Y) and the pixel value at place, α determines according to virtual view position, and scope is between 0 ~ 1;

Cavity noise processed is as follows:

I _I(X _I,Y)＝ω ₁I _IL(X _I,Y)+(1-ω ₁)I _IR(X _I,Y)

I _IL(X _I,Y)＝I _IL(X _L+(1-α)*d ^IL,Y)＝I _L(X _L,Y)

ω ¹represent virtual view position and left and right view location proportionate relationship, X _l+ (1-α) * d ^iLfor virtual view position and left view position relationship, I _iL(X _i, Y) and be (X by left view transformation to intermediate virtual viewpoint pixel coordinate _i, Y) and the pixel value at place, d ^rLfor the parallax value of left and right view, I _l(X _l, Y) for left view pixel coordinate be (X _l, Y) and the pixel value at place, I _i(X _i, Y) for final virtual view be (X at pixel coordinate _i, Y) and the pixel value at place.

As the further preferred version of a kind of view synthesizing method based on mean shift Stereo matching of the present invention, the conditional plan in described step 1.3 refers to that spatial domain is less than h _s, color gamut is less than h _r;

Wherein, h _sspatial bandwidth parameter, h _rit is color bandwidth parameter.

The present invention adopts above technical scheme compared with prior art, has following technique effect:

1, view synthesizing method provided by the invention, draw for virtual view and have good effect, drawing view quality is relatively high.

2, the view of the inventive method synthesis has good image effect, and clear-cut is clearly demarcated, has good stereoeffect.

Accompanying drawing explanation

Fig. 1 is method flow diagram of the present invention;

Fig. 1 (a) is list window disparity map of the present invention;

Fig. 1 (b) is multiwindow disparity map of the present invention;

Fig. 2 (a) is the left view of the multiwindow stereo matching method based on adaptive weight and the weights multiwindow stereo matching method contrast effect figure based on mean shift Region Segmentation;

Fig. 2 (b) is the normal view of the multiwindow stereo matching method based on adaptive weight and the weights multiwindow stereo matching method contrast effect figure based on mean shift Region Segmentation;

Fig. 2 (c) is the disparity map obtained based on the weights multiwindow stereo matching method of mean shift Region Segmentation;

Fig. 2 (d) is the disparity map obtained based on the multiwindow stereo matching method of adaptive weight;

Fig. 3 (a) is the left view that algorithm of the present invention obtains;

Fig. 3 (b) is the intermediate virtual view that algorithm of the present invention obtains;

Fig. 3 (c) is the right view that algorithm of the present invention obtains.

Embodiment

Below with reference to specific embodiment, technical scheme provided by the invention is described in detail, following embodiment should be understood and be only not used in for illustration of the present invention and limit the scope of the invention.

As shown in Figure 1, a kind of view synthesizing method based on mean shift Stereo matching, specifically comprises the steps:

Step 1, adopts mean shift method to carry out Region Segmentation to original image;

Step 1.1, is converted to the chrominance space matched with image by image;

Step 1.2, utilizes the smoothing operation of mean shift and then obtains the convergence point of each pixel;

Step 1.3, merges the convergence point of each pixel according to the conditional plan preset;

Step 2, adopts the multiwindow stereo matching method based on adaptive weight to obtain disparity map corresponding to stereo-picture;

Step 2.1, certain pixel constructs one and supports window in a reference image;

Step 2.2, moves in parallel this support window, carrys out the similitude between calculation window with similarity measure function along disparity range; Specifically be calculated as follows:

$C_{SAD}(x,y)=\underset{c\∈\left\{r,g,b\right\}}{\Σ}|I_c\left(x\right)-I_c\left(y\right)|$

$\begin{array}{c}{C}_{GRAD}(x,y)=\underset{c\∈\left\{r,g,b\right\}}{\Σ}{\left|{\▿}_x{I}_c\left(x\right)-{\▿}_x{I}_c\left(y\right)\right|}^2+\\ \underset{c\∈\left\{r,g,b\right\}}{\Σ}{\left|{\▿}_y{I}_c\left(x\right)-{\▿}_y{I}_c\left(y\right)\right|}^2\end{array}$

C(x,y)＝(1-w)*C _GRAD(x,y)+w*C _SAD(x,y)

Wherein, C _sAD(x, y) is for asking for the absolute value sum of current pixel point and field pixel r, g, b triple channel color distortion, I _cx () is current pixel point value of color, x is current point coordinate, I _c(y) for field value of color around current point, y be world coordinates, with the horizontal and vertical gradient of representative image respectively, w is the weights between 0 and 1, C _gRAD(x, y) for ask for current point and its field pixel in the horizontal direction with the absolute value sum of vertical direction r, g, b triple channel color distortion, C (x, y) is similarity measure function.

Step 3, to the smoothing operation of step 2 gained disparity map;

Step 4, carries out forward view interpolation and empty noise processed to disparity map after smooth operation, and then completes Image Rendering;

Step 4.1, obtains positional information and the colouring information of each pixel;

Step 4.2, then by the disparity map determination virtual view position relationship of this view, determine evolution relation, with correspondence former depending on colouring information color filling is carried out to virtual view, forward view interpolation formula is as follows:

I _IR(X _I,Y)＝I _IR(X _R+(1-α)*d ^RL,Y)＝I _R(X _R,Y)

Wherein, X _r+ (1-α) * d ^rLfor virtual view position and right view position relationship, I _iR(X _i, Y) and for transform to intermediate virtual viewpoint pixel coordinate by right view be (X _i, Y) and the pixel value at place, d ^rLfor the parallax value of left and right view, I _r(X _r, Y) for right view pixel coordinate be (X _r, Y) and the pixel value at place, α determines according to virtual view position, and scope is between 0 ~ 1; Cavity noise processed is as follows:

I _I(X _I,Y)＝ω ₁I _IL(X _I,Y)+(1-ω ₁)I _IR(X _I,Y)

I _IL(X _I,Y)＝I _IL(X _L+(1-α)*d ^IL,Y)＝I _L(X _L,Y)

ω ¹represent virtual view position and left and right view location proportionate relationship, X _l+ (1-α) * d ^iLfor virtual view position and left view position relationship, I _iL(X _i, Y) and be (X by left view transformation to intermediate virtual viewpoint pixel coordinate _i, Y) and the pixel value at place, d ^rLfor the parallax value of left and right view, I _l(X _l, Y) for left view pixel coordinate be (X _l, Y) and the pixel value at place, I _i(X _i, Y) for final virtual view be (X at pixel coordinate _i, Y) and the pixel value at place.

Specific as follows: step 1, adopt meanshift method to carry out Region Segmentation to image:

One width figure ties up color vector by a two-dimensional spatial location coordinate and M and forms, and as M=1, this image is gray-scale map, and as M=3, this image is RGB figure, and as M>3, this image is multivariate joint probability image.Each pixel of image is determined by colouring information and spatial positional information, and so the characteristic vector of mean shift filter is defined as X=[x ^s, x ^r] ^t.In setting image, certain pixel space position is x, and being the center of circle with x, take h as radius, drops on the some position x in higher-dimension ball _idefine two pattern rules:

The color of 1.x pixel and x _ipixel color is more close, and it is higher that we define probability density;

2. from x position more close to pixel x _i, definition probability density is higher.

Aforesaid probability density function is:

$K_{h_{s,}{h}_r}\left(x\right)=\frac{c}{h_s^2{h}_r^2}K\left(\left|\right|\frac{x^s-x_i^s}{h_s}|{|}^2\right)K\left(\left|\right|\frac{x^r-x_i^r}{h_r}|{|}^2\right)$

Wherein: represent spatial positional information, decentre pixel is far away, and its value is larger; representative color information, color is more close, and its value is larger; x ^slocus coordinate, x ^rcolor of image feature, it is the kernel function of Two Variables; h _sspatial bandwidth parameter, control area s _hsize, be the threshold parameter of central space distance; h _rbe color bandwidth parameter, being class pixel value spacing parameter, is to spatial domain s _hselect again; C is normaliztion constant; Number of pixels parameter M in another setting class, merges into a class when pixel value is less than this value in single class, be equivalent to the once selection between class.

Introduce gaussian kernel function in Mean Shift algorithm, range of application is region S _hsize, and Iamge Segmentation generally uses gaussian kernel function.So Mean Shift vector expression is:

$y_{j+1}=\frac{\underset{i=1}{\overset{n}{\Σ}}G\left(\left|\right|\frac{x_i-y_j}{h}\left|\right|\right)x_i}{\underset{i=1}{\overset{n}{\Σ}}G\left(\left|\right|\frac{x_i-y_j}{h}\left|\right|\right)},j=1,2,\mathrm{...}$

y _j+1-y _j＝M _h(y _j)

Make x _iand z _i(i=1,2,3...n) represents that d ties up original graph picture point and convergence point respectively, L _ifor the region labeling of i-th after Iamge Segmentation.Based on above-mentioned vector form, we utilize Mean Shift algorithm to carry out Region Segmentation, specifically comprise the steps:

Step 1-1, is converted to suitable chrominance space by image.Cromogram need be converted to L*U*V chrominance space and operate, and gray-scale map directly can carry out next step;

Step 1-2, utilizes the smoothing operation of mean shift:

A) j=1 is made, y _{i, 1}=x _i;

B) current pixel central point y is calculated _{i, j+1};

C) mean shift vector m is calculated _h,G(y _i,j)=y _{i, j+1}-y _i,j;

D) above-mentioned operation is repeated until find convergence point

Step 1-3, by the convergence point z of each pixel _imerge according to certain conditional plan, the conditional plan in this example refers to that spatial domain is less than h _s, color gamut is less than h _r;

Step 2, adopts the multiwindow stereo matching method based on adaptive weight to obtain disparity map corresponding to stereo-picture:

This step have employed a kind of local matching algorithm, and assume that in match window, all pixels are on a similar depth plane, these pixels have similar parallax value.The basic thought of the method can obtain accurate matching result in degree of depth discontinuity zone and the same area, will carry out Stereo matching for each pixel chooses suitable match window adaptively.It is crucial that the determination of similarity measure function in Stereo matching process, when asking for initial parallax figure, certain pixel constructs one and supports window in a reference image, this window is moved in parallel along disparity range, the similitude between calculation window is carried out with similarity measure function, the consistency namely calculated between pixel reaches maximum, and error energy function is minimum.

Fig. 1 (a) is list window disparity map of the present invention; Fig. 1 (b) is multiwindow disparity map of the present invention;

Specifically, this step comprises the steps:

Step 2.1, certain pixel constructs one and supports window in a reference image;

Step 2.2, moves in parallel this window along disparity range, carry out the similitude between calculation window with similarity measure function, and the consistency namely calculated between pixel reaches maximum, and error energy function is minimum.

That the similarity measure function in this example adopts is the Matching power flow function C (x that absolute difference sum SAD (sum of absolute difference) and image gradient information combine, y), as following formula (3) ~ (5);

$C_{SAD}(x,y)=\underset{c\∈\left\{r,g,b\right\}}{\Σ}\left|I_c\left(x\right)-I_c\left(y\right)\right|$

$\begin{array}{c}{C}_{GRAD}(x,y)=\underset{c\∈\left\{r,g,b\right\}}{\Σ}{\left|{\▿}_x{I}_c\left(x\right)-{\▿}_x{I}_c\left(y\right)\right|}^2+\\ \underset{c\∈\left\{r,g,b\right\}}{\Σ}{\left|{\▿}_y{I}_c\left(x\right)-{\▿}_y{I}_c\left(y\right)\right|}^2\end{array}$

C(x,y)＝(1-w)*C _GRAD(x,y)+w*C _SAD(x,y)

Wherein with the horizontal and vertical gradient of representative image respectively, w is the weights between 0 and 1.

This cost function is that the color similarity of pixel in correlation window and distance similarity are combined consideration, correlation window adopts the method for multiwindow to make each pixel matching to best pixel point, design a cost function based on weights, in change window, accumulate error energy function.

Color similarity weights formula:

$f_s\left({\mathrm{\Δc}}_{pq}\right)=\exp (-\frac{{\mathrm{\Δc}}_{pq}}{{\mathrm{\γ}}_c})$

Wherein, Δ c _pqit is the color distortion of two pixels; γ _cfor evaluate color Similarity Parameter, in this example, get 3.

Δ c in above formula _pq=| R _p-R _q|+| G _p-G _q|+| B _p-B _q| (7).

Multiwindow weight computing formula in reference-view and target view:

$w^{\′}(p,q)=\left\{\begin{array}{c}1,q\∈S_p\\ \exp (-\frac{{\mathrm{\Δc}}_{pq}}{{\mathrm{\γ}}_c}),q\∈S_p\end{array}\right.$

Wherein, S _pit is a p moving area.

Figure segmentation weighted cumulative error energy function: $E^{\′}(p,{\stackrel{-}{p}}_d)=\frac{{\Σ}_{q\∈N_{\mathrm{\ρ}},{\stackrel{-}{q}}_d\∈N_{\stackrel{-}{\mathrm{\ρ}},d}}{w}^{\′}(p,q)w^{\′}({\stackrel{-}{p}}_d,{\stackrel{-}{q}}_d)C(q,{\stackrel{-}{q}}_d)}{{\Σ}_{q\∈N_{\mathrm{\ρ}},{\stackrel{-}{q}}_d\∈N_{\stackrel{-}{\mathrm{\ρ}},d}}{w}^{\′}(p,q)w^{\′}({\stackrel{-}{p}}_d,{\stackrel{-}{q}}_d)}$

Wherein, the length of M and N presentation video and width, d' is disparity estimation result, and d is standard disparity map.When error energy function is minimum, the disparity map that stereo-picture is corresponding can be obtained.

In order to prove the superiority of this step matching process, introducing error hiding amount e herein and weighing disparity estimation effect:

$e=\underset{i=1}{\overset{M}{\Σ}}\underset{j=1}{\overset{N}{\Σ}}|d^{\′}(i,j)-d(i,j)|$

Wherein, the length of M and N presentation video and width, d' is disparity estimation result, and d is standard disparity map.Fig. 1 is the disparity map based on the weights multiwindow of mean shift Region Segmentation and the left view of single Window match.The algorithm that the present invention proposes is drawn for virtual view good effect, and drawing view quality is relatively high.Fig. 2 (a) is the left view of the multiwindow stereo matching method based on adaptive weight and the weights multiwindow stereo matching method contrast effect figure based on mean shift Region Segmentation; Fig. 2 (b) is the normal view of the multiwindow stereo matching method based on adaptive weight and the weights multiwindow stereo matching method contrast effect figure based on meanshift Region Segmentation; Fig. 2 (c) is the disparity map obtained based on the weights multiwindow stereo matching method of meanshift Region Segmentation; Fig. 2 (d) is the disparity map obtained based on the multiwindow stereo matching method of adaptive weight;

Next, we compare with two groups of experiment stereo-pictures the weights multiwindow stereo matching method based on mean shift Region Segmentation adopted based on the multiwindow stereo matching method of adaptive weight and the present invention:, split plot design e=15029998 adaptive method e=15045738.Obviously, the image segmentation that the image that the method that the present invention adopts obtains is good, less view error, clear-cut, has good stereoeffect.

Step 3, adopts forward interpolation method to draw view, specifically comprises the steps:

Step 3-1, first travels through each pixel in former view, obtains positional information and the colouring information of each pixel;

Step 3-2, then by the disparity map determination virtual view position relationship of this view, determine evolution relation, with correspondence former depending on colouring information color filling is carried out to virtual view, interpolation formula is as follows:

I _IR(X _I,Y)＝I _IR(X _R+(1-α)*d ^RL,Y)＝I _R(X _R,Y)

Wherein, α determines according to virtual view position, and scope is between 0 ~ 1.

In step 2, cannot reappear for some borders in view and hidden place, these somes parallax value in disparity map is 0, the cavitation namely occurred.In order to address this problem, first determine the position of cavity point, the virtual view utilizing left view to go out at virtual location carries out colouring information filling to some position, cavity:

I _IL(X _I,Y)＝I _IL(X _L+(1-α)*d ^IL,Y)＝I _L(X _L,Y)

I _I(X _I,Y)＝I _IL(X _I,Y)+I _IR(X _I,Y)

We verify effect of the present invention under Matlab2012b software environment, in emulation quarry at present to generally acknowledge test patterns such as " Tsukuba " that field Middlebury dataset provides to and the standard disparity map of each image.The medial view adopting algorithm herein to obtain and the contrast schematic diagram of left and right reference-view, Fig. 3 (a) is the left view that algorithm of the present invention obtains;

Fig. 3 (b) is the intermediate virtual view that algorithm of the present invention obtains; Fig. 3 (c) is the right view that algorithm of the present invention obtains., simulation result shows, adopt the view of the inventive method synthesis to have good image effect, clear-cut is clearly demarcated, has good stereoeffect.

The present invention proposes a kind of new virtual view method for drafting.First the two kinds of gradation of image images and coloured image commonly used in life are considered, mean shift image segmentation is adopted to carry out Region Segmentation to known reference image, recycle the weights multiwindow matching method based on the color similarity between pixel, and introduce Matching power flow function, optimize initial parallax value on this basis thus obtain the disparity map of more accurate stereo pairs, then to the smoothing operation of disparity map; Finally carry out forward view interpolation and empty noise processed, finally complete the Image Rendering of optional position between the view of known left and right.There is good stereoeffect.

, and for gray-scale map and cromogram all applicable.

Those skilled in the art of the present technique are understandable that, unless otherwise defined, all terms used herein (comprising technical term and scientific terminology) have the meaning identical with the general understanding of the those of ordinary skill in field belonging to the present invention.Should also be understood that those terms defined in such as general dictionary should be understood to have the meaning consistent with the meaning in the context of prior art, unless and define as here, can not explain by idealized or too formal implication.

Above embodiment is only and technological thought of the present invention is described, can not limit protection scope of the present invention with this, and every technological thought proposed according to the present invention, any change that technical scheme basis is done, all falls within scope.By reference to the accompanying drawings embodiments of the present invention are explained in detail above, but the present invention is not limited to above-mentioned execution mode, in the ken that those of ordinary skill in the art possess, makes a variety of changes under can also or else departing from the prerequisite of present inventive concept.

Claims (5)

1. based on a view synthesizing method for mean shift Stereo matching, it is characterized in that, specifically comprise the steps:

Step 1, adopts mean shift method to carry out Region Segmentation to original image;

Step 2, adopts the multiwindow stereo matching method based on adaptive weight to obtain disparity map corresponding to stereo-picture;

Step 3, to the smoothing operation of step 2 gained disparity map;

Step 4, carries out forward view interpolation and empty noise processed to disparity map after smooth operation, and then completes Image Rendering.

2. a kind of view synthesizing method based on mean shift Stereo matching according to claim 1, is characterized in that: described step 1 detailed process is as follows:

Step 1.1, is converted to the chrominance space matched with image by image;

Step 1.2, utilizes the smoothing operation of mean shift and then obtains the convergence point of each pixel;

Step 1.3, merges the convergence point of each pixel according to the conditional plan preset.

3. a kind of view synthesizing method based on mean shift Stereo matching according to claim 1, is characterized in that: described step 2 detailed process is as follows:

Step 2.1, certain pixel constructs one and supports window in a reference image;

Step 2.2, moves in parallel this support window, carrys out the similitude between calculation window with similarity measure function along disparity range; Specifically be calculated as follows:

C(x,y)＝(1-w)*C _GRAD(x,y)+w*C _SAD(x,y)

Wherein, C _sAD(x, y) for asking for the absolute value sum of current pixel point and field pixel r, g, b triple channel color distortion,

I _cx () is current pixel point value of color, x is current point coordinate, I _c(y) for field value of color around current pixel point, y be world coordinates,

with the horizontal and vertical gradient of representative image respectively, w is the weights between 0 and 1,

C _gRAD(x, y) for ask for current point and its field pixel in the horizontal direction with the absolute value sum of vertical direction r, g, b triple channel color distortion, C (x, y) is similarity measure function.

4. a kind of view synthesizing method based on mean shift Stereo matching according to claim 1, is characterized in that: described step 4 detailed process is as follows:

Step 4.1, obtains positional information and the colouring information of each pixel;

Step 4.2, then by the disparity map determination virtual view position relationship of this view, determine evolution relation, with correspondence former depending on colouring information color filling is carried out to virtual view, forward view interpolation formula is as follows:

I _IR(X _I,Y)＝I _IR(X _R+(1-α)*d ^RL,Y)＝I _R(X _R,Y)

Wherein, X _r+ (1-α) * d ^rLfor virtual view position and right view position relationship, I _iR(X _i, Y) and for transform to intermediate virtual viewpoint pixel coordinate by right view be (X _i, Y) and the pixel value at place, d ^rLfor the parallax value of left and right view, I _r(X _r, Y) for right view pixel coordinate be (X _r, Y) and the pixel value at place, α determines according to virtual view position, and scope is between 0 ~ 1;

Cavity noise processed is as follows:

I _I(X _I,Y)＝ω ₁I _IL(X _I,Y)+(1-ω ₁)I _IR(X _I,Y)

I _IL(X _I,Y)＝I _IL(X _L+(1-α)*d ^IL,Y)＝I _L(X _L,Y)

ω ₁represent virtual view position and left and right view location proportionate relationship, X _l+ (1-α) * d ^iLfor virtual view position and left view position relationship, I _iL(X _i, Y) and be (X by left view transformation to intermediate virtual viewpoint pixel coordinate _i, Y) and the pixel value at place, d ^rLfor the parallax value of left and right view, I _l(X _l, Y) for left view pixel coordinate be (X _l, Y) and the pixel value at place, I _i(X _i, Y) for final virtual view be (X at pixel coordinate _i, Y) and the pixel value at place.

5. a kind of view synthesizing method based on mean shift Stereo matching according to claim 2, it is characterized in that, the conditional plan in described step 1.3 refers to that spatial domain is less than h _s, color gamut is less than h _r;

Wherein, h _sspatial bandwidth parameter, h _rit is color bandwidth parameter.