CN102708569B - Based on the monocular infrared image depth estimation method of SVM model - Google Patents

Abstract

The present invention proposes a kind of method of carrying out estimation of Depth for infrared image, utilizes support vector regression theoretical to building estimation of Depth model to the depth characteristic that infrared image extracts, thus utilizes depth model to estimate the depth map of new infrared image.First the proper vector in infrared image is chosen; Then utilize gradually linear regression method to search the proper vector relevant to infrared depth characteristic, again independent component analysis is carried out to the proper vector of screening, draw mutually independently proper vector, and build infrared degree of depth training set; Then based on the non-linear support vector regression theory with kernel function, estimation of Depth model is calculated to training set; By model, estimation of Depth is carried out to it after finally the new infrared image introduced being extracted according to the feature extracting method in training set, thus draw the depth map of this infrared image.The results show, the method effectively can estimate the depth map of infrared image.

Description

Based on the monocular infrared image depth estimation method of SVM model

Technical field

The present invention relates to a kind of method can carrying out estimation of Depth to monocular infrared image, can be estimated the spatial positional information of the scene in infrared image by this method.

Background technology

Namely the estimation of Depth of image is obtain depth distance information from image, is the problem of a depth perception in essence.The relative distance on the surface detected during what the spatial positional information built by depth perception characterized is from observer to scene.The algorithm that the depth distance information in coloured image recovered is now existing more satisfactory, but for infrared image, because of its reflection is the Temperature Distribution of scene, and have the defect such as low signal-to-noise ratio, low contrast, the degree of depth algorithm recovering this image still belongs to blank.If the depth information of infrared image can be recovered, so will greatly improve the understanding effect of human eye to this image.

Current image depth estimation method is mainly for binocular depth clue with launch based on the estimation of Depth of image sequence, and these two kinds of methods all depend on the feature difference between image.And monocular depth is estimated, comparatively classical in the middle of algorithm traditional is in early days " shape from shade (shape from shading) ", this algorithm with space multistory geometry for theoretical foundation, the light and shade change produced to body surface according to the light source irradiation i.e. shade of image to recover Object Depth, but needs priori (as reflection model and light source direction etc.) to make the limitation increase of application because of this algorithm.Afterwards, some researchers find the importance of experience gradually, start to utilize the method for machine learning to go to address this problem.The team of Stanford University Andrew Ng carries out estimation of Depth by the model utilizing Markov field to train to single image, reaches good effect; The Aloysha Efros team of CMU is simple classification in manual Calibration Field scape before training then, such as sky, trees, ground and perpendicular line etc., then utilize a large amount of data to learn these classifications, and new images classified eventually through structure Bayesian model thus recovers depth information.Although this method is comparatively applicable to the simple picture of a series of scene, and produces a desired effect, inaccurate often for the classification not having in scene to learn.

Summary of the invention

The object of this invention is to provide a kind of method of carrying out estimation of Depth for infrared image, more adequately can be estimated the depth information of infrared image by the method.

In order to achieve the above object, technical scheme of the present invention there is provided a kind of monocular infrared image depth estimation method based on SVM model, and it is characterized in that, step is:

Step 1, obtain any monocular infrared image I (x, y) and with the depth map corresponding to this monocular infrared image I (x, y);

Step 2, for monocular infrared image I (x, y) characteristic area on each pixel setting at least three different scales in, three different scales are ascending to be respectively: the first yardstick, second yardstick and the 3rd yardstick, wherein, the characteristic area of each yardstick at least comprises the image block that is positioned at center and with on this image block, under, left, right adjacent image block, the image block that be positioned at center of i-th pixel on the first yardstick is i-th pixel itself, the image block that be positioned at center of this i-th pixel on the second yardstick comprises all image blocks on the first yardstick, the image block that be positioned at center of this i-th pixel on the 3rd yardstick comprises all image blocks on the second yardstick, by that analogy,

Step 3, calculate monocular infrared image I (x, the proper vector of the characteristic area y) corresponding to each pixel, the characteristic component of the proper vector of i-th pixel at least comprises: the gray-scale value of all image blocks of i-th pixel on the first yardstick, the texture energy of each image block of i-th pixel on other yardsticks except the first yardstick, the gradient energy of different directions of each image block of i-th pixel on other yardsticks except the first yardstick and the average of all gradient energy and variance, and the sharpness of each image block of i-th pixel on other yardsticks except the first yardstick,

Step 4, utilize the method for gradually linear regression analysis and independent component analysis to screen successively to all proper vectors obtained in step 3, comparatively met the proper vector of infrared image depth information;

Step 5, utilize by step 4 screening obtain proper vector and this monocular infrared image I (x, y) depth map of described correspondence builds the set of degree of depth training sample, by the maps feature vectors in the set of degree of depth training sample in another new feature space, and the depth value support vector regression of new feature space and depth map is carried out nonlinear fitting, and then build depth model;

Step 6, the depth model analysis that the new monocular infrared image collected is obtained by step 5 is obtained estimation of Depth value.

Preferably, texture energy described in step 3 is calculated by Louth mask, and concrete steps are:

Adopt N number of two-dimentional Louth mask substantially, be designated as M ₁..., M _n, described monocular infrared image I (x, y) and each two-dimentional Louth mask are done convolution, then the value after monocular infrared image I (x, y) and a kth two-dimentional Louth mask convolution is: T _k(x, y)=I (x, y) × M _k, k=1 ..., N, then i-th m the image block N of pixel on jth yardstick ⁱm texture energy that () obtains after monocular infrared image I (x, y) with a kth two-dimentional Louth mask convolution is

${\mathrm{En}}_{N^i\left(m\right)}^j=\underset{x,y\∈N^i\left(m\right)}{\mathrm{\Σ}}\left|T_k(x,y)\right|.$

Preferably, the average of gradient energy described in step 3, gradient energy and the calculation procedure of variance are:

On x-axis direction and y-axis direction, x-axis gradient map I is tried to achieve respectively to described monocular infrared image I (x, y) _gradx(x, y) and y-axis gradient map I _grady(x, y), then monocular infrared image I (x, y) is at angle θ _lgrad G on direction _l(x, y)=I _gradx(x, y) × cos (θ _l)+I _grady(x, y) × sin (θ _l), wherein, l=0 ..., (L-1), L is total number in direction, calculates the gradient energy of the different directions of each image block of each pixel on other yardsticks except the first yardstick subsequently, wherein, i-th m the image block N of pixel on jth yardstick ⁱm () is at angle θ _lgradient energy on direction is then i-th m the image block N of pixel on jth yardstick ⁱm () is at angle θ _lthe average of the gradient energy on direction i-th m the image block N of pixel on jth yardstick ⁱm () is at angle θ _lgradient energy variance on direction wherein, size is image block N ⁱthe number of m pixel that () comprises,

Preferably, in described step 4, the concrete steps of gradually linear regression analysis are:

In step 4.1.1, calculating proper vector, the related coefficient of each characteristic component and depth value, obtains the sequence of each characteristic component to effect of depth degree according to the absolute value of related coefficient is descending;

Step 4.1.2, from the characteristic component of the maximum absolute value of related coefficient, progressively introduce regression equation, and do regression equation significance test, if significantly can not think, selected all characteristic components are not all the principal elements of influence depth value, if introduce regression equation one by one successively from depth value impact is descending significantly again;

Step 4.1.3, the characteristic component that often introducing one is new all need to carry out significance test to each characteristic component contained in regression equation, by that in new regression equation not significantly and minimum characteristic component rejection is affected on depth value, repeat this step until each characteristic component in regression equation is remarkable;

The characteristic component had the greatest impact to depth value in step 4.1.4, the characteristic component do not introduced again, repeats step 4.1.3 and step 4.1.4, until cannot reject selected characteristic component, till also cannot introducing new characteristic component.

Preferably, in described step 4, the concrete steps of independent component analysis are:

Step 4.2.1, specify the proper vector of M pixel after gradually linear regression analysis to be observation data X, and carry out centralization to observation data X, making it average is 0;

Step 4.2.2, by the observation data X albefaction of centralization, after projecting to new subspace by observation data X, become albefaction vector Z, Z=W ₀x, wherein, W ₀for whitening matrix, W ₀=Λ ^-1/2u ^t, Λ is the eigenvalue matrix of the covariance matrix of observation data X, and U is the eigenvectors matrix of the covariance matrix of observation data X;

Step 4.2.3, renewal W ^*make W ^*=E{Zg (W ^tz) }-E{g ' (W ^tz) } W, wherein g () is nonlinear function, and then to W ^*standardization: W=W ^*/ || W ^*||, if do not restrain, repeat this step, wherein, select a mould to be that the initial random weight vector of 1 is as the initial value of W.

The present invention proposes a kind of monocular infrared image depth estimation method based on SVM model, the method utilizes infrared image " spatial context " and " multiple dimensioned " information extraction depth characteristic vector, by the method for gradually linear regression analysis and independent component analysis, the feature extracted is screened successively, be conducive to finding the depth characteristic being more applicable to infrared image, and build degree of depth training set with this; Adopting support vector regression theoretical to training set to carry out nonlinear fitting, improve the accuracy rate of matching.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of a kind of monocular infrared image depth estimation method based on SVM model provided by the invention.

Embodiment

For making the present invention become apparent, hereby with a preferred embodiment, and accompanying drawing is coordinated to be described in detail below.

As shown in Figure 1, present embodiment discloses a kind of monocular infrared image depth estimation method based on SVM model, the steps include:

Step 1, obtain any monocular infrared image I (x, y) and with the depth map corresponding to this monocular infrared image I (x, y);

Step 2, for monocular infrared image I (x, y) each pixel in sets the characteristic area on three different scales, three different scales are ascending to be respectively: the first yardstick, second yardstick and the 3rd yardstick, wherein, the characteristic area of each yardstick comprises the image block that is positioned at center and with on this image block, under, left, right adjacent image block, the image block that be positioned at center of i-th pixel on the first yardstick is i-th pixel itself, the image block that be positioned at center of this i-th pixel on the second yardstick comprises all image blocks on the first yardstick, the image block that be positioned at center of this i-th pixel on the 3rd yardstick comprises all image blocks on the second yardstick,

Step 3, calculate monocular infrared image I (x, the proper vector of the characteristic area y) corresponding to each pixel, the characteristic component of the proper vector of i-th pixel at least comprises: the gray-scale value of all image blocks of i-th pixel on the first yardstick, the texture energy of each image block of i-th pixel on the second yardstick and the 3rd yardstick, gradient energy on 8 directions of each image block of i-th pixel on the second yardstick and the 3rd yardstick and the average of all gradient energy and variance, and the sharpness of each image block of i-th pixel on the second yardstick and the 3rd yardstick,

The texture energy of each image block of i-th pixel on the second yardstick and the 3rd yardstick is calculated by following steps:

Adopt the two-dimentional Louth mask that 9 basic, be designated as M ₁..., M ₉, monocular infrared image I (x, y) and each two-dimentional Louth mask are done convolution, then the value after monocular infrared image I (x, y) and a kth two-dimentional Louth mask convolution is: T _k(x, y)=I (x, y) × M _k, k=1 ..., 9, then i-th m the image block N of pixel on jth yardstick ⁱm texture energy that () obtains after monocular infrared image I (x, y) with a kth two-dimentional Louth mask convolution is ${\mathrm{En}}_{N^i\left(m\right)}^j=\underset{x,y\∈N^i\left(m\right)}{\mathrm{\Σ}}\left|T_k(x,y)\right|,$ j＝2，3。

The gradient energy of different directions of each image block of i-th pixel on the second yardstick and the 3rd yardstick and the average of all gradient energy and variance are calculated by following steps:

On x-axis direction and y-axis direction, x-axis gradient map I is tried to achieve respectively to monocular infrared image I (x, y) _gradx(x, y) and y-axis gradient map I _grady(x, y), then monocular infrared image I (x, y) is at angle θ _lgrad G on direction _i(x, y)=I _gradx(x, y) × cos (θ _l)+I _grady(x, y) × sin (θ _l), wherein, l=0 ..., 7, calculate the gradient energy of the different directions of each image block of each pixel on other yardsticks except the first yardstick subsequently, wherein, i-th m the image block N of pixel on jth yardstick ⁱm () is at angle θ _lgradient energy on direction is then i-th m the image block N of pixel on jth yardstick ⁱm () is at angle θ _lthe average of the gradient energy on direction i-th m the image block N of pixel on jth yardstick ⁱm () is at angle θ _lgradient energy variance on direction wherein, size is image block N ⁱthe number of m pixel that () comprises,

The sharpness of each image block of i-th pixel on the second yardstick and the 3rd yardstick is calculated by following steps:

I-th m the image block N of pixel on jth yardstick ⁱthe sharpness of (m) wherein, I (x, y) is image block N ⁱ(m) containing the gray scale of pixel, for image block N ⁱthe average of the gray scale of (m) contained pixel,

Step 4, utilize the method for gradually linear regression analysis and independent component analysis to screen successively to all proper vectors obtained in step 3, comparatively met the proper vector of infrared image depth information.

The concrete steps of gradually linear regression analysis in step 4 are:

In step 4.1.1, calculating proper vector, the related coefficient of each characteristic component and depth value, obtains the sequence of each characteristic component to effect of depth degree according to the absolute value of related coefficient is descending;

Step 4.1.2, from the characteristic component of the maximum absolute value of related coefficient, progressively introduce regression equation, and do regression equation significance test, if significantly can not think, selected all characteristic components are not all the principal elements of influence depth value, if introduce regression equation one by one successively from depth value impact is descending significantly again;

Step 4.1.3, the characteristic component that often introducing one is new all need to carry out significance test to each characteristic component contained in regression equation, by that in new regression equation not significantly and minimum characteristic component rejection is affected on depth value, repeat this step until each characteristic component in regression equation is remarkable;

The characteristic component had the greatest impact to depth value in step 4.1.4, the characteristic component do not introduced again, repeats step 4.1.3 and step 4.1.4, until cannot reject selected characteristic component, till also cannot introducing new characteristic component.

In step 4, independent component analysis is Fast ICA algorithm, by Fast ICA algorithm to analyzing through the proper vector of gradually linear regression analysis, makes between each component independent as much as possible.In the present embodiment, Fast ICA algorithm is maximum as search direction using negentropy, and its concrete steps are as follows:

Step 4.2.1, specify the proper vector of M pixel after gradually linear regression analysis to be observation data X, and carry out centralization to observation data X, making it average is 0;

Step 4.2.2, by the observation data X albefaction of centralization, after projecting to new subspace by observation data X, become albefaction vector Z, Z=W ₀x, wherein, W ₀for whitening matrix, W ₀=Λ ^-1/2u ^t, Λ is the eigenvalue matrix of the covariance matrix of observation data X, and U is the eigenvectors matrix of the covariance matrix of observation data X;

Step 4.2.3, renewal W ^*make W ^*=E{Zg (W ^tz) }-E{g ' (W ^tz) } W, wherein g () is nonlinear function, and then to W ^*standardization: W=W ^*/ || W ^*||, if do not restrain, repeat this step, wherein, select a mould to be that the initial random weight vector of 1 is as the initial value of W.

The depth map of step 5, the proper vector utilizing the infrared image drawn by step 4 and infrared image builds degree of depth training sample set { f _i, depth _i, wherein, f _ibe the proper vector of i-th pixel, i=1 ..., l, f _i∈ χ, depth _ifor the depth value corresponding to i-th pixel.Then utilize support vector regression the Theory Construction depth model, and then estimation of Depth is carried out to new infrared image.Concrete steps are as follows:

Step 5.1, proper vector f _ifirst be mapped to some new feature space F by a certain mapping phi: χ → F, and the relation of training sample set be expressed as form:

${\mathrm{depth}}_i=W_d^T\mathrm{\φ}\left(f_i\right)+b_d;$

Wherein, need by the regression coefficient vector solved, b _d∈ R is unknown deviator.B _dbelong to real number space, can be regarded as the nonlinear fitting of the input space here, by proper vector f _ithe input space at place is mapped to feature space (higher dimensional space), and solve the regression coefficient vector of feature space, the result that matching draws remains real number space, namely with b _dand depth _iaffiliated space consistent.

Step 5.2, to solve namely be requirement minimum, also can be equivalent to and solve following optimization problem:

$\underset{W_d^T,b_d,\mathrm{\ξ},{\mathrm{\ξ}}^{*}}{\min }\frac{1}{2}{W}_d^T{W}_d+C_d\underset{i=1}{\overset{l}{\mathrm{\Σ}}}({\mathrm{\ξ}}_i+{\mathrm{\ξ}}_i^{*}),$

$s.t.:W_d^T\mathrm{\Φ}\left(f_i\right)+b_d-{\mathrm{depth}}_i\≤\mathrm{\ϵ}+{\mathrm{\ξ}}_i,$

${{\mathrm{depth}}_i-W}_d^T\mathrm{\Φ}\left(f_i\right)-b_d\≤\mathrm{\ϵ}+{\mathrm{\ξ}}_i^{*},$

${\mathrm{\ξ}}_i,{\mathrm{\ξ}}_i^{*}\≥0,i=1,...,l.$

Wherein ε > 0 is biased error, ξ _iwith be respectively the bound of lax item, C _d> 0 is penalty factor, is the constant that certain is specified.

Step 5.3, by above formula substitute into structure Lagrangian function obtain its dual form:

$\underset{\mathrm{\α},{\mathrm{\α}}^{*}}{\min }\frac{1}{2}{(\mathrm{\α}-{\mathrm{\α}}^{*})}^{T}Q(\mathrm{\α}-{\mathrm{\α}}^{*})+\mathrm{\ϵ}\underset{i=1}{\overset{l}{\mathrm{\Σ}}}({\mathrm{\α}}_i+{\mathrm{\α}}_i^{*})+\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{depth}}_i({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*}),$

$s.t.:\underset{i=1}{\overset{l}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*})=\mathrm{0,0}\≤{\mathrm{\α}}_i,{\mathrm{\α}}_i^{*}\≤C_d,i=1,...l.$

Its Kernel Function Q _ij=K (f _i, f _j)=Φ (f _i) ^tΦ (f _j), utilize Novel Algorithm can be in the hope of optimum solution α and α ^*, substitute in the formula of step 5.1 relational expression namely obtained between the proper vector of i-th pixel and its depth value, be expressed as follows:

${\mathrm{depth}}_i=\underset{i=1}{\overset{l}{\mathrm{\Σ}}}(-{\mathrm{\α}}_i+{\mathrm{\α}}_i^{*})K(f_i,f)+b_d.$

Step 5.4, to new infrared image by obtaining the proper vector of each pixel after the feature extraction of step 1 and step 2 and screening, then obtain the depth value corresponding with it according to the formulae discovery of step 5.1, just complete the estimation to new infrared image depth value.

Claims (5)

1., based on a monocular infrared image depth estimation method for SVM model, it is characterized in that, step is:

Step 1, obtain any monocular infrared image I (x, y) and with the depth map corresponding to this monocular infrared image I (x, y);

Step 2, for monocular infrared image I (x, y) characteristic area on each pixel setting at least three different scales in, three different scales are ascending to be respectively: the first yardstick, second yardstick and the 3rd yardstick, wherein, the characteristic area of each yardstick at least comprises the image block that is positioned at center and with on this image block, under, left, right adjacent image block, the image block that be positioned at center of i-th pixel on the first yardstick is i-th pixel itself, the image block that be positioned at center of this i-th pixel on the second yardstick comprises all image blocks on the first yardstick, the image block that be positioned at center of this i-th pixel on the 3rd yardstick comprises all image blocks on the second yardstick, by that analogy,

Step 3, calculate monocular infrared image I (x, the proper vector of the characteristic area y) corresponding to each pixel, the characteristic component of the proper vector of i-th pixel at least comprises: the gray-scale value of all image blocks of i-th pixel on the first yardstick, the texture energy of each image block of i-th pixel on other yardsticks except the first yardstick, the gradient energy of different directions of each image block of i-th pixel on other yardsticks except the first yardstick and the average of all gradient energy and variance, and the sharpness of each image block of i-th pixel on other yardsticks except the first yardstick,

Step 4, utilize the method for gradually linear regression analysis and independent component analysis to screen successively to all proper vectors obtained in step 3, comparatively met the proper vector f of infrared image depth information _i;

Step 5, utilize by step 4 screening obtain proper vector f _iwith this monocular infrared image I (x, y) depth map of described correspondence builds the set of degree of depth training sample, by the maps feature vectors in the set of degree of depth training sample in another new feature space, and the depth value support vector regression of new feature space and depth map is carried out nonlinear fitting, and then structure depth model, obtain estimation of Depth value by depth model analysis, the steps include:

Step 5.1, proper vector f _ifirst mapped φ: χ → F is mapped to new feature space F, and by the relation depth of training sample set _ibe expressed as form:

${\mathrm{depth}}_i=W_d^T\mathrm{\φ}\left(f_i\right)+b_d;$

Wherein, need by the regression coefficient vector solved, b _d∈ R is unknown deviator, b _dbelong to real number space, regard the nonlinear fitting of the input space here as, by proper vector f _ithe input space at place is mapped to high-dimensional feature space, solves the regression coefficient vector of feature space, and the result that matching draws remains real number space, namely with b _dand depth _iaffiliated space consistent;

Step 5.2, to solve namely be requirement minimum, also can be equivalent to and solve following optimization problem:

$\underset{W_d^T,b_d,\mathrm{\ξ},{\mathrm{\ξ}}^{*}}{\min }\frac{1}{2}{W}_d^T{W}_d+C_d\underset{i=1}{\overset{l}{\mathrm{\Σ}}}({\mathrm{\ξ}}_i+{\mathrm{\ξ}}_i^{*}),$

$s.t.:W_d^T\mathrm{\φ}\left(f_i\right)+b_d-{\mathrm{depth}}_i\≤\mathrm{\ϵ}+{\mathrm{\ξ}}_i,$

${\mathrm{depth}}_i-W_d^T\mathrm{\φ}\left(f_i\right)-b_d\≤\mathrm{\ϵ}+{\mathrm{\ξ}}_i^{*},$

${\mathrm{\ξ}}_i,{\mathrm{\ξ}}_i^{*}\≥0,i=1,...,l,$

Wherein ε > 0 is biased error, ξ _iwith be respectively the bound of lax item, C _d> 0 is penalty factor, is the constant of specifying;

Step 5.3, by step 5.2 formula substitute into structure Lagrangian function obtain its dual form:

$\underset{\mathrm{\α},{\mathrm{\α}}^{*}}{\min }\frac{1}{2}{(\mathrm{\α}-{\mathrm{\α}}^{*})}^{T}Q(\mathrm{\α}-{\mathrm{\α}}^{*})+\mathrm{\ϵ}\underset{i=1}{\overset{l}{\mathrm{\Σ}}}({\mathrm{\α}}_i+{\mathrm{\α}}_i^{*})+\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{depth}}_i({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*}),$

$s.t.:\underset{i=1}{\overset{l}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*})=\mathrm{0,0}\≤{\mathrm{\α}}_i,{\mathrm{\α}}_i^{*}\≤C_d,i=1,...,l,$

Its Kernel Function Q=K (f _i, f _j)=φ (f _i) ^tφ (f _j), utilize Novel Algorithm can be in the hope of optimum solution α and α ^*, substitute in the formula of step 5.1 relational expression namely obtained between the proper vector of i-th pixel and its depth value, be expressed as follows:

${\mathrm{depth}}_i=\underset{i=1}{\overset{l}{\mathrm{\Σ}}}(-{\mathrm{\α}}_i+{\mathrm{\α}}_i^{*})K(f_i,f)+b_d;$

Step 5.4, to new infrared image by obtaining the proper vector of each pixel after the feature extraction of step 3 and step 4 and screening, then obtain the depth value corresponding with it according to the formulae discovery of step 5.3, just complete the estimation to new infrared image depth value.

2. a kind of monocular infrared image depth estimation method based on SVM model as claimed in claim 1, it is characterized in that, texture energy described in step 3 is calculated by Louth mask, and concrete steps are:

Adopt N number of two-dimentional Louth mask substantially, be designated as M ₁..., M _n, described monocular infrared image I (x, y) and each two-dimentional Louth mask are done convolution, then the value after monocular infrared image I (x, y) and a kth two-dimentional Louth mask convolution is: T _k(x, y)=I (x, y) × M _k, k=1 ..., N, then i-th m the image block N of pixel on jth yardstick ⁱm texture energy that () obtains after monocular infrared image I (x, y) with a kth two-dimentional Louth mask convolution is ${\mathrm{En}}_{N^i\left(m\right)}^j=\underset{x,y\∈N^i\left(m\right)}{\mathrm{\Σ}}\left|T_k(x,y)\right|.$

3. a kind of monocular infrared image depth estimation method based on SVM model as claimed in claim 1, it is characterized in that, the average of gradient energy described in step 3, gradient energy and the calculation procedure of variance are:

On x-axis direction and y-axis direction, x-axis gradient map I is tried to achieve respectively to described monocular infrared image I (x, y) _gradx(x, y) and y-axis gradient map I _grady(x, y), then monocular infrared image I (x, y) is at angle θ _lgrad G on direction _l(x, y)=I _gradx(x, y) × cos (θ _l)+I _grady(x, y) × sin (θ _l), wherein, l=0 ..., (L-1), L is total number in direction, calculates the gradient energy of the different directions of each image block of each pixel on other yardsticks except the first yardstick subsequently, wherein, i-th m the image block N of pixel on jth yardstick ⁱm () is at angle θ _lgradient energy on direction is then i-th m the image block N of pixel on jth yardstick ⁱm () is at angle θ _lthe average of the gradient energy on direction i-th m the image block N of pixel on jth yardstick ⁱm () is at angle θ _lgradient energy variance on direction wherein, size is image block N ⁱthe number of m pixel that () comprises,

4. a kind of monocular infrared image depth estimation method based on SVM model as claimed in claim 1, it is characterized in that, in described step 4, the concrete steps of gradually linear regression analysis are:

In step 4.1.1, calculating proper vector, the related coefficient of each characteristic component and degree of depth position, obtains the sequence of each characteristic component to effect of depth degree according to the absolute value of related coefficient is descending;

Step 4.1.2, from the characteristic component of the maximum absolute value of related coefficient, progressively introduce regression equation, and do regression equation significance test, if significantly can not think, selected all characteristic components are not all the principal elements of influence depth value, if introduce regression equation one by one successively from depth value impact is descending significantly again;

Step 4.1.3, the characteristic component that often introducing one is new all need to carry out significance test to each characteristic component contained in regression equation, by that in new regression equation not significantly and minimum characteristic component rejection is affected on depth value, repeat this step until each characteristic component in regression equation is remarkable;

The characteristic component had the greatest impact to depth value in step 4.1.4, the characteristic component do not introduced again, repeats step 4.1.3 and step 4.1.4, until cannot reject selected characteristic component, till also cannot introducing new characteristic component.

5. a kind of monocular infrared image depth estimation method based on SVM model as claimed in claim 1, it is characterized in that, in described step 4, the concrete steps of independent component analysis are:

Step 4.2.1, specify the proper vector of M pixel after gradually linear regression analysis to be observation data X, and carry out centralization to observation data X, making it average is 0;

Step 4.2.2, by the observation data X albefaction of centralization, after projecting to new subspace by observation data X, become albefaction vector Z, Z=W ₀x, wherein, W ₀for whitening matrix, W ₀=Λ ^-1/2u ^t, Λ is the eigenvalue matrix of the covariance matrix of observation data X, and U is the eigenvectors matrix of the covariance matrix of observation data X;

Step 4.2.3, renewal W ^*make W ^*=E{Zg (W ^tz) }-E{g ' (W ^tz) } W, wherein g () is nonlinear function, and then to W ^*standardization: W=W ^*/ || W ^*||, if do not restrain, repeat this step, wherein, select a mould to be that the initial random weight vector of 1 is as the initial value of W.