CN102609936A - Stereo image matching method based on belief propagation - Google Patents

Abstract

The invention relates to a stereo image matching method based on belief propagation, which comprises the following steps of: a, carrying out Rank transform for an input image, and then causing the input image to be transformed into a Rank image; b, optimizing the smooth item of the image through a global energy function; c, calculating the minimum of the global energy function according to a belief propagation algorithm; d, obtaining a parallax value of pixels according to the minimum of belief; e, and using the parallax value as a gray value corresponding to the pixels, and outputting a parallax image. The method achieves the purpose of obtaining parallax by calculating the minimum of the global energy function. In the calculating process, the Rank transform of the image is added, so that the constraint condition among neighborhood pixels is enhanced, the smooth item is improved, therefore the noise immunity and image smoothness in the belief propagation process are increased, and the precision of calculation is increased.

Description

Image solid matching method based on confidence spread

Technical field

The present invention relates to image solid matching method based on confidence spread.

Background technology

Human human vision system is very perfect, a complicated sensory perceptual system through the extraneous world of visually-perceptible, and the efficient of visually-perceptible environment is very high, and the external information more than 80% obtains through vision.Along with the development of infotech, let the human visual performance of machine simulation become possibility.After signal processing technology and computing machine occurred, people attempted to obtain ambient image and convert thereof into digital signal with video camera, with the overall process of computer realization to Vision information processing, like this, have just formed a new subject: computer vision.The goal in research of computer vision is to make computing machine have the ability through the cognitive three-dimensional environment information of two dimensional image; So that computing machine can be experienced the environment in the visual field; Understand the content of perception; Comprise shape, position, attitude motion etc., and can they be described, store, discern and understand, and make behavior or decision-making on this basis.

Three-dimensional coupling is the important content in computer vision research field, also is a step the most key and difficult in the stereoscopic vision.Its final goal is the disparity map that obtains matching image.Disparity map is to be benchmark with the arbitrary width of cloth image of image pair, and its size is the size of this benchmark image, and element value is the image of parallax value.Parallax is exactly to observe the direction difference that same target produced from two points that certain distance is arranged.Three-dimensional matching problem can be expressed as the minimization problem of innings energy of demanding perfection usually, at first defines energy function, through various optimization methods, asks its minimum value afterwards.

Confidence spread algorithm in the three-dimensional coupling is based on the Stereo Matching Algorithm of markov random file (MRF).Markov random file is comprising the two-layer meaning, and one is Markov character, and one is random field.Markov character is meant that when a sequence of random variables arranged by the order of time successively, it had no relation in N+1 distribution character and the stochastic variable before n-hour constantly.The problem that possesses this character just meets Markov character.Random field be meant in giving each position according to certain distribute give a value of phase space at random after, it all just is referred to as random field.

In the confidence spread algorithm, adopt the actual grey value of pixel to carry out the solid coupling at present, actual effect is often inaccurate, has therefore also influenced the correctness of final judgement.

Summary of the invention

To the problem of above appearance, the invention provides a kind of image solid matching method based on confidence spread, strengthen the constraint between neighborhood territory pixel, improve the noise immunity and the image smoothing property of confidence spread process, to improve the precision of Stereo Matching Algorithm.

The present invention is based on the image solid matching method of confidence spread, comprising:

A. the image with input carries out converting the Rank image into after the Rank conversion;

B. be optimized through the level and smooth item of global energy function image;

C. according to confidence spread algorithm computation global energy minimum of a function value;

D. obtain the parallax value of pixel according to the minimum value of degree of confidence;

E. with the gray-scale value of described parallax value, export anaglyph as corresponding pixel points.

The Rank conversion is a kind of common method in the Flame Image Process; Be to be benchmark with the center pixel in the rectangular window; Do comparison with the gray-scale value of other pixel in the window; Statistics is represented the gray-scale value of benchmark pixel less than the number of benchmark pixel (being the center pixel in the window) gray-scale value with this number.Method of the present invention is not to carry out the three-dimensional coupling of image with the actual grey value of pixel; But the sequence number that adopts the gray-scale value ordering of benchmark pixel in the neighborhood window is calculated number+1 of benchmark pixel gray-scale value (in the ordering sequence number=window of center pixel gray scale less than); So just strengthen the constraint between neighborhood territory pixel, improved the noise immunity of three-dimensional coupling; When setting up the global energy function, level and smooth item is optimized, improves treatment effect,, improved the precision of Stereo Matching Algorithm through this improvement of two to the image border.

Concrete, the described Rank conversion of step a comprises:

A1. in the window of predetermined size, by from left to right with from the gray-scale value of the directional statistics window interior pixel that falls down;

A2. replace the gray-scale value of center pixel according to the ordering sequence number of center pixel gray scale in the window;

A3. the gray-scale value of each pixel of entire image is all used with the relative gray-scale value of center pixel and represented.

Preferably, learn that the window through 3 * 3 pixel sizes calculates can obtain result more accurately through lot of test.

Further, among the step c through to parallax coupling similarity, image level and smooth and adjacent iterations the time the said global energy minimum of a function of the parallax information anded value of different pixels point.

Concrete, the iterative step to said confidence spread algorithm among the step c comprises:

C1. information initializing, it is that average probability distributes that the nodal information that transmits between different pixels point is unified assignment;

C2. from 1 to the nodal information the said different pixels point of setting threshold cycle and regeneration of mature;

C3. the degree of confidence of calculating pixel point;

C4. the maximal value of calculating the degree of confidence that obtains is the maximum a posteriori probability of pixel.Described posterior probability refers in Bayesian statistics, and a random occurrence is a resulting conditional probability after considering relevant evidence or data.

Concrete, the described threshold value of step c2 is 50～100.

Test learns that method of the present invention is to reach the purpose of trying to achieve parallax through asking for global energy minimum of a function value.In the computation process, add the Rank conversion of image, strengthened the constraint condition between neighborhood territory pixel point thus, improved level and smooth, thereby improved the noise immunity and the image smoothing property of confidence spread process, improved the degree of accuracy of calculating.

Below in conjunction with the embodiment of embodiment, foregoing of the present invention is remake further detailed description.But should this be interpreted as that the scope of the above-mentioned theme of the present invention only limits to following instance.Do not breaking away under the above-mentioned technological thought situation of the present invention, various replacements or change according to ordinary skill knowledge and customary means are made all should comprise within the scope of the invention.

Embodiment

The present invention is based on the image solid matching method of confidence spread, comprising:

A. the image with input carries out converting the Rank image into after the Rank conversion, comprises step:

A1. in the window of 3 * 3 pixels, by from left to right with from the gray-scale value of the directional statistics window interior pixel that falls down;

A2. replace the gray-scale value of center pixel according to the ordering sequence number of center pixel gray scale in the window;

A3. the gray-scale value of each pixel of entire image is all used with the relative gray-scale value of center pixel and represented;

B. calculate the global energy function:

$E\left(d\right)=\underset{p\∈P}{\mathrm{\Σ}}{D}_p\left(d_p\right)+\underset{(p,q)\∈N}{\mathrm{\Σ}}V(d_p,d_q),$ Wherein

$V(d_i,d_j)=\left\{\begin{array}{cc}0& d_i=d_j\\ {\mathrm{\ρ}}_I\left(\mathrm{\ΔI}\right)& \mathrm{other}\end{array}\right.$

${\mathrm{\&rho;}}_I\left(\mathrm{\&Delta;I}\right)=\left\{\begin{array}{cc}P\&times;|d_i-d_j|& \mathrm{\&Delta;I}<T\\ P^{\&prime;}\&times;|d_i-d_j|& \mathrm{other}\end{array}\right.$

Variable declaration:

E (d): global energy function.

D _p(d _p): d _pRepresent the locational parallax of this p, D _p(d _p) be coupling similarity measure function.

P ∈ P:P representes the set of all pixels, and the p remarked pixel belongs to the pixel in the P set.

V (d _p, d _q): level and smooth function, the normally difference functions of two neighbor parallaxes.

N: the set of neighbor.

V (d _i, d _j): level and smooth function, the normally difference functions of two neighbor parallaxes.

d _j: j point pixel parallax value.

d _i: i point pixel parallax value.

ρ _I(Δ I): the gradient delta I function between neighborhood territory pixel.

P ': the parameter of setting.

T: preset threshold.

Through global energy function E (d) level and smooth of image is optimized;

C. the said global energy minimum of a function of the parallax information anded value of different pixels point through to parallax coupling similarity, image level and smooth and adjacent iterations the time:

$m_{p\&RightArrow;q}^t\left(d_q\right)=\underset{d_p\&Element;\mathrm{\&Omega;}}{\min }(D_p\left(d_p\right)+V(d_p,d_q)+\underset{s\&Element;N\left(p\right)\backslash q}{\mathrm{\&Sigma;}}{m}_{s\&RightArrow;p}^{t-1}\left(d_p\right)).$

Wherein:

During

the t time iteration; Pixel p passes to the parallax information of pixel q, and t is since 1.

d _q: d _pRepresent the locational parallax of this p.

Ω: the parallax hunting zone of remarked pixel point.

S: the s time iteration.

N (p) q: remove the p neighborhood of a point pixel outside the q point.

During

the t-1 time iteration, pixel s passes to the parallax information of pixel p.

In the iterative computation therein, node x _iPass to x _jMessage be designated as m _Ij(x _j), i is node x _jThe neighborhood territory pixel number, value is 8, observer nodes y _iPass to and treat observer nodes x _iMessage be designated as m _i(x _i), x _jDegree of confidence use b _j(x _j) expression.m _Ij(x _j), m _j(x _j) and b _j(x _j) this three is the L dimensional vector, L representes x _iThe number of possible value.Therefore the step of iteration comprises:

C1. information initializing is with nodal information m _Ij(x _j) unified assignment is that average probability distributes, i.e. m _Ij(x _j)=1/L, and m _j(x _j, y _j)=φ (x _j, y _j), φ (x _j, y _j) be local evidence;

C2. from the nodal information 1 to the setting threshold T said different pixels point of cycle and regeneration of mature, wherein the value of threshold value T is 50～100:

$m_{\mathrm{ij}}^{t+1}\left(x_j\right)\&LeftArrow;\underset{x_i}{\max }{\mathrm{\&Psi;}}_{\mathrm{ij}}(x_i,x_j)m_i^t\left(x_i\right)\underset{x_k\&Element;N\left(x_i\right)\backslash x_k}{\mathrm{\&Pi;}}{m}_{\mathrm{ki}}^t\left(x_i\right);$

Wherein:

T is since 1, node x during t+1 iteration _iPass to x _jMessage.

←: sign passes to the value on " ← " the right the variable on " ← " left side.

Ψ _Ij(x _i, x _j): node x _iAnd x _jBetween compatible dependence.

During the t time iteration, x _jNeighborhood N (x _i) remove x in the scope _jBehind the point, the product of nodal information.

C3. the degree of confidence of calculating pixel point:

$b_i\left(x_i\right)\&LeftArrow;m_i\left(x_i\right)\underset{x_k\&Element;N\left(x_i\right)}{\mathrm{\&Pi;}}{m}_{\mathrm{ki}}\left(x_i\right);$

C4. the maximal value of calculating the degree of confidence that obtains is pixel x _iMaximum a posteriori probability:

$x_i^{\mathrm{MAP}}=\underset{x_i}{\mathrm{Arg}\mathrm{Max}}{b}_i\left(x_i\right),$ b _i(x _i) be node x _iDegree of confidence;

D. obtain the parallax value of pixel according to the minimum value of degree of confidence:

$d_p^{*}=\underset{d_p\&Element;\mathrm{\&Omega;}}{\arg \min }{b}_p\left(d_p\right).$

Wherein:

The parallax value of the p point position that

finally obtains.

b _p(d _p): the degree of confidence of pixel p;

E. with the gray-scale value of described parallax value, export anaglyph as corresponding pixel points.

Claims (6)

1. based on the image solid matching method of confidence spread, its characteristic comprises:

A. the image with input carries out converting the Rank image into after the Rank conversion;

B. be optimized through the level and smooth item of global energy function image;

C. according to confidence spread algorithm computation global energy minimum of a function value;

D. obtain the parallax value of pixel according to the minimum value of degree of confidence;

E. with the gray-scale value of described parallax value, export anaglyph as corresponding pixel points.

2. the image solid matching method based on confidence spread as claimed in claim 1 is characterized by: the described Rank conversion of step a comprises:

A1. in the window of predetermined size, by from left to right with from the gray-scale value of the directional statistics window interior pixel that falls down;

A2. replace the gray-scale value of center pixel according to the ordering sequence number of center pixel gray scale in the window;

A3. the gray-scale value of each pixel of entire image is all used with the relative gray-scale value of center pixel and represented.

3. the image solid matching method based on confidence spread as claimed in claim 2 is characterized by: described window is the window of 3 * 3 pixels.

4. the image solid matching method based on confidence spread as claimed in claim 1 is characterized by: among the step c through to parallax coupling similarity, image level and smooth and adjacent iterations the time the said global energy minimum of a function of the parallax information anded value of different pixels point.

5. like the described image solid matching method based on confidence spread of one of claim 1 to 4, it is characterized by: the iterative step to said confidence spread algorithm among the step c comprises:

C1. information initializing, it is that average probability distributes that the nodal information that transmits between different pixels point is unified assignment;

C2. from 1 to the nodal information the said different pixels point of setting threshold cycle and regeneration of mature;

C3. the degree of confidence of calculating pixel point;

C4. the maximal value of calculating the degree of confidence that obtains is the maximum a posteriori probability of pixel.

6. the image solid matching method based on confidence spread as claimed in claim 5 is characterized by: the described threshold value of step c2 is 50～100.