CN106780297A - Image high registration accuracy method under scene and Varying Illumination - Google Patents

Abstract

The present invention relates to a kind of image high registration accuracy method under scene and Varying Illumination, step includes：Camera reorientation, the initialization of L and F, illumination correction, the correction of camera geometric position.The present invention proposes multiple dimensioned low-rank conspicuousness detection method, and by with the collaboration conspicuousness priori based on GMM, the detection of multiple dimensioned low-rank conspicuousness is generalized in multiple image collaboration conspicuousness detection, to detect the same or analogous region occurred in multiple image.Compared with traditional conspicuousness detection method, the multiple dimensioned super-pixel blending algorithm of low-rank proposed by the invention solves the problems, such as that scale selection is difficult, and achieves more reliable conspicuousness testing result.

Description

Image high registration accuracy method under scene and Varying Illumination

Technical field

The invention belongs to image registration field, it is related to a kind of towards matching somebody with somebody to high level image high accuracy under scene and illumination variation Quasi- method.The rapid relocation technology of camera is applied to the image high accuracy that scene and illumination all changes and matched somebody with somebody by this method In quasi- problem, with the 1 principal direction image occurred with last time observed image in camera repositioning process and 6 auxiliary directional images pair Present image high registration accuracy is target.

Background technology

The background technology being related in the present invention has：

(1) change detection (Change Detection)：Change detection is the important preprocessing step of high-rise vision application. In detection steps and decision rule in change detection, CDNet provides an extensive REF video data set for reality, And good ranking is kept in change detection algorithm.Many algorithms based on most recently newly proposition, such as SOBS, SC_SOBS, Sub SENSE etc., the background modeling in CDNet is one of most successful strategy.The development that other are worth mentioning is included based on change Change the three-dimensional voxel (3Dvoxel) of detection and changed with the urban scale structure of range data using multiple panorama sketch and visit Survey.In order to solve the influence of illumination variation and camera motion, it is aobvious with luminosity that current classic method is conceived to space exploration The change of work, and the implicit delicate change that must be processed as noise.However, occurring in space and light varience not at some On important small yardstick, the hypothesis on this basis largely limits its ability for delicate texture variations detection. Our invention shows that low rank analysis (low-rank analysis) can be used to decompose the sparse change in several pictures.

(2) color continuity (Color Constancy)：In change detection, quick color continuity extensively should Use correction illumination variation.Due to the requirement of real-time speed, most of change detection methods can only be connected with simple static state color Continuous property treating method, so limits the ability of the frequent and violent illumination variation of its tolerance.In addition, intrinsic image can For the difference of different illumination under correction Same Scene.Recently, the decomposition of intrinsic image has been further extended various Part, including the shape of decomposition, illumination, reflectivity etc. from one or plurality of pictures.But, these nearest development Or needing complicated majorized function, or needs to capture plurality of pictures under intensive lighting condition, this causes them not Inexpensive, end-to-end, close-grained change detection can be directly applied for.

(3) geometric correction (Geometry Correction)：Geometric correction is that another is visited in the change with robustness Indispensable part in survey.Conventional method for rigid scene includes similitude, affine or projective transformation.For non- Rigid dynamic scene, light stream can be used to correct the deviation of the camera solid having dislocation.Many photographs are considered in our invention The condition of express contract beam, employs the SIFT streams of extension to carry out geometric correction.

The content of the invention

Accurately detecting, will to the deficiency of finegrained image change in high level scene for prior art for the present invention Camera reorientation is applied in fine change detection, by obtaining scene, illumination variation under image, there is provided it is a kind of end-to-end The new technology for ensureing change detection robustness and accuracy from coarse to fine, is conducive to further improving to there is delicate fine texture to become Change detectivity.

Therefore, the present invention is adopted the following technical scheme that, mainly comprise the following steps：

1. camera reorientation：

Collect the scene observation sample of different time.7 pictures are gathered in one scenario：1 ambient lighting and 6 are fixed Illuminated to side.These pictures are stored as column vector.

Given X is as follows as above-mentioned 7 the constituted matrixes of illumination：

Wherein, 1≤k≤K, K are the number (taking K=6 here) of side illumination orientations.

Current camera is reoriented to and observes close angles and positions with last time, by observed with last time multiple DSL is adjusted roughly after comparing, and obtains current observing matrix Y：

Camera relocates step：

A () initializes Current camera to rational position so that this position is big enough in comprising real target Region.I_cRepresent posture and the position of photo current correspondence camera；

B () keeps a rectangle frame R for blueness_bIn I_cCenter, a navigation rectangle frame R for red_rRepresent Current camera Posture is poor with the relative geometry of target area.Posture and the position of dynamic adjustment camera are until there is following equation to set up：

R_r=HR_b

Wherein H is by I_cAnd X_ELThe unit matrix for calculating.

The initialization of 2.L (illumination difference) and F (camera geometric correction stream)：

L represents illumination difference, and F represents the geometric correction stream of camera.

By X, Y for being given in (1), it is assumed that there is image x_iAnd y_i：Wherein, i=0 interval scales EL images, 1≤i≤K interval scales DSL images.Global linear photometric calibration matrix is obtainedWith an offset vectorFormula：

Wherein,WithDifference representative image x_iAnd y_iMatching SIFT feature RGB color.(*) is defined specification illumination Model, can solve under the form of closing.

3. it is normal to perceive illumination correction：

Based on Lambertian reflection model (Lambertian Reflectance Model)：

I_p=∫ ＜ n_p, w ＞ ρ_pL(w)dw

Wherein, I_p、n_pAnd ρ_pColor, normality and the albedo of pixel p are represented respectively, and L (w) is grading function.

By to X_FMiddle addition virtual optical L^v() corrects its illumination deviation with Current observation Y, obtains equation below：

Wherein,It is the color after pixel p correction,It is to add the color increment produced after virtual luminosity.

Therefore, the luminosity for balancing X and Y with following function is poor, function minimization that will be below：

Wherein, LⁱIt is that the spatial variations of X and Y under i-th illumination normally perceive illumination difference.The first half of formula is encouragedWith yⁱLuminosity uniformity.It is variable, wherein 0≤C_p≤ 1 is the possibility of pixel p change.After formula Half part encourages the continuity in virtual optical space, w_pqRepresent the similitude of p and q, p~q and represent that p is adjacent with q.

4. camera geometric position correction：

Illumination difference L is perceived according to normalⁱ, a new luminosity X being corrected can be obtained_L.By extending SIFT frames Frame, corrects its energy function as follows：

Wherein,It is identical with (3).

So far, by Pictures location, the illumination that can obtain to above-mentioned steps being shot with last time almost complete identical Picture, can carry out detection and find its fine and closely woven texture variations.

Proposed by the invention merges to carry out conspicuousness and cooperate with what conspicuousness was detected by the multiple dimensioned super-pixel of low-rank Technology is mainly included the following steps that：

1) detection of the conspicuousness of single scale；

2) fusion of multiple dimensioned conspicuousness；

3) refinement of conspicuousness；

The present invention achieves more reliable conspicuousness testing result, helps further to improve prior art to conspicuousness inspection The disposal ability of survey.

Advantages and positive effects of the present invention：

The present invention proposes multiple dimensioned low-rank conspicuousness detection method, and by with the collaboration conspicuousness elder generation based on GMM Test, the detection of multiple dimensioned low-rank conspicuousness is generalized in multiple image collaboration conspicuousness detection, occur to detect in multiple image Same or analogous region.Compared with traditional conspicuousness detection method, the multiple dimensioned super-pixel of low-rank proposed by the invention Blending algorithm solves the problems, such as that scale selection is difficult, and achieves more reliable conspicuousness testing result.

Brief description of the drawings

Fig. 1：Image high registration accuracy flow chart

Fig. 2：Summer Palace figure of buddha slight change detection figure

Fig. 3：Camera resets bitmap

Specific embodiment

The invention will be further described below in conjunction with the accompanying drawings and by specific embodiment, and following examples are descriptive , it is not limited, it is impossible to which protection scope of the present invention is limited with this.

A kind of image high registration accuracy method under scene and Varying Illumination, step is as follows：

1. camera reorientation：

Collect the scene observation sample of different time.7 pictures are gathered in one scenario：1 ambient lighting and 6 are fixed Illuminated to side.These pictures are stored as column vector.

Given X is as follows as above-mentioned 7 the constituted matrixes of illumination：

Wherein, 1≤k≤K, K are the number (taking K=6 here) of side illumination orientations.

Current camera is reoriented to and observes close angles and positions with last time, by observed with last time multiple DSL is adjusted roughly after comparing, and obtains current observing matrix Y：

Camera relocates step：

A () initializes Current camera to rational position so that this position is big enough in comprising real target Region.I_cRepresent posture and the position of photo current correspondence camera；

B () keeps a rectangle frame R for blueness_bIn I_cCenter, a navigation rectangle frame R for red_rRepresent Current camera Posture is poor with the relative geometry of target area.Posture and the position of dynamic adjustment camera are until there is following equation to set up：

R_r=HR_b

Wherein H is by I_cAnd X_ELThe unit matrix for calculating.

The initialization of 2.L (illumination difference) and F (camera geometric correction stream)：

L represents illumination difference, and F represents the geometric correction stream of camera.

By X, Y for being given in (1), it is assumed that there is image x_iAnd y_i：Wherein, i=0 interval scales EL images, 1≤i≤K interval scales DSL images.Global linear photometric calibration matrix is obtainedWith an offset vectorFormula：

Wherein,WithDifference representative image x_iAnd y_iMatching SIFT feature RGB color.(*) is defined specification illumination Model, can solve under the form of closing.

3. it is normal to perceive illumination correction：

Based on Lambertian reflection model (Lambertian Reflectance Model)：

I_p=∫ ＜ n_p, w ＞ ρ_pL(w)dw

Wherein, I_p、n_pAnd ρ_pColor, normality and the albedo of pixel p are represented respectively, and L (w) is grading function.

By to X_FMiddle addition virtual optical L^v() corrects its illumination deviation with Current observation Y, obtains equation below：

Wherein,It is the color after pixel p correction,It is to add the color increment produced after virtual luminosity.

Therefore, the luminosity for balancing X and Y with following function is poor, function minimization that will be below：

Wherein, LⁱIt is that the spatial variations of X and Y under i-th illumination normally perceive illumination difference.The first half of formula is encouragedWith yⁱLuminosity uniformity.It is variable, wherein 0≤C_p≤ 1 is the possibility of pixel p change.After formula Half part encourages the continuity in virtual optical space, w_pqRepresent the similitude of p and q, p~q and represent that p is adjacent with q.

4. camera geometric position correction：

Illumination difference L is perceived according to normalⁱ, a new luminosity X being corrected can be obtained_L.By extending SIFT frames Frame, corrects its energy function as follows：

Wherein,It is identical with (3).

The image significance detection method of low-rank Multiscale Fusion is to be based on low-rank Multiscale Fusion method, will be registered Two good images, with it, so as to the conspicuousness for different zones is detected.

Its operating procedure is as follows：

Step S1：Prepare two images of reference picture and present image, and be loaded into the method.

Step S2：Present image is carried out the initialization of illumination, light stream, is processed by present image and reference picture, Current state hypograph is processed into image of the last state hypograph same light according under, it is preliminary eliminate before and after image irradiation twice Influence.

Step S3：Current state image is carried out into illumination correction, last state hypograph to current state figure below is calculated The transformation matrix of picture, by present image plus change moment matrix, is finally reached the effect of normal illumination correction.

Step S4：Present image is carried out into geometric correction, two side-play amounts of image is calculated, and side-play amount is applied to In present image, the effect of camera geometric correction is finally realized.

Claims (2)

1. a kind of image high registration accuracy method under scene and Varying Illumination, step is as follows：

(1) the scene observation sample of different time is collected, N pictures are gathered in one scenario：1 ambient lighting and N-1 are individual fixed Illuminated to side, these pictures are stored as column vector,

Given X is as follows as the constituted matrix of above-mentioned N number of illumination：

$X=\[x_{EL},x_{{\mathrm{DSL}}_1},...,x_{{\mathrm{DSL}}_k}\]$

Wherein, 1≤k≤K, K are the number of side illumination orientations,

Current camera is reoriented to and observed close angles and positions with last time, by multiple the DSL ratios observed with last time To rear rough adjustment, current observing matrix Y is obtained：

$Y=\[y_{EL},y_{{\mathrm{DSL}}_1},...,y_{{\mathrm{DSL}}_k}\]$

(2) the initialization of L and F：

L represents illumination difference, and F represents the geometric correction stream of camera,

By X, Y for being given in (1), it is assumed that there is image x_iAnd y_i：Wherein, i=0 interval scales EL images, 1≤i≤K interval scales DSL figures Picture, has obtained global linear photometric calibration matrixWith an offset vectorFormula：

$\begin{array}{cc}\[{\hat{A}}^i,{\hat{b}}^i\]=\arg \underset{A^i,b^i}{min}\left|\right|{\hat{A}}^i{\tilde{x}}^i+b^i-{\tilde{y}}^i|{|}_F^2,& (*)\end{array}$

Wherein,WithDifference representative image x_iAnd y_iMatching SIFT feature RGB color, (*) is defined specification illumination mould Type, can solve under the form of closing；

(3) it is normal to perceive illumination correction：

Based on Lambertian reflection model：

I_p=∫ ＜ n_p, w ＞ ρ_pL(w)dw

Wherein, I_p、n_pAnd ρ_pColor, normality and the albedo of pixel p are represented respectively, and L (w) is grading function,

By to X_FMiddle addition virtual optical L^v() corrects its illumination deviation with Current observation Y, obtains equation below：

$X_{{\mathrm{LF}}_p}=\&Integral;<n_p,w>{\mathrm{\&rho;}}_p\left(L^x\right(w)+L^v(w\left)\right)dw,$

$X_{{\mathrm{LF}}_p}=x_{F_p}+L_p^v=y_p,$

Wherein,It is the color after pixel p correction,It is to add the color increment produced after virtual luminosity,

The luminosity for balancing X and Y with following function is poor, function minimization that will be below：

$L^i=\arg \underset{L^v}{min}{\mathrm{\&Sigma;}}_p{(x_{F_p}^i+L_p^v-y_p^i)}^2\exp (-\frac{C_p}{\mathrm{\&sigma;}})+{\mathrm{\&alpha;\&Sigma;}}_{p~q}{w}_{pq}{(L_p^v-L_q^v)}^2$

Wherein, LⁱIt is that the spatial variations of X and Y under i-th illumination normally perceive illumination difference, the first half of formula is encouragedWith yⁱLuminosity uniformity,It is variable, wherein 0≤C_p≤ 1 be pixel p change possibility, formula it is latter half of Divide the continuity for encouraging virtual optical space, w_pqRepresent the similitude of p and q, p~q and represent that p is adjacent with q,

(4) camera geometric position correction：

Illumination difference L is perceived according to normalⁱ, a new luminosity X being corrected can be obtained_L, by extending SIFT frameworks, repair Just its energy function is as follows：

$E\left(F\right)=\underset{i,p}{\&Sigma;}\left|\right|x_L^i(p+F_p)-y^i\left(p\right)|{|}_1\exp (-\frac{C_p}{\mathrm{\&sigma;}})+{\mathrm{\&beta;\&Sigma;}}_p|\left|F_p\right|{|}_2^2+{\mathrm{\&Sigma;}}_{p~q}min\left(\mathrm{\&gamma;}\right||F_p-F_q|{|}_1,d)$

Wherein,It is identical with (3),

Pictures location, the identical picture of illumination for obtaining being shot with last time by above-mentioned steps, carry out detection and find that its is thin Close texture variations.

2. the image high registration accuracy method under scene according to claim 1 and Varying Illumination, it is characterised in that： Described camera reorientation step：

A () initializes Current camera to rational position so that this position is big enough in comprising real target area, I_cRepresent posture and the position of photo current correspondence camera；

B () keeps a rectangle frame R for blueness_bIn I_cCenter, a navigation rectangle frame R for red_rRepresent Current camera posture Relative geometry with target area is poor, and posture and the position of dynamic adjustment camera are until there is following equation to set up：

R_r=HR_b

Wherein H is by I_cAnd X_ELThe unit matrix for calculating.