CN104616298A - Method for detecting moving target of ink-jet printing fabric based on mixed-state Gauss MRF (Markov Random Field) model - Google Patents

Abstract

A method for detecting a moving target of an ink-jet printing fabric based on a mixed-state Gauss MRF model comprises the steps: (1) inputting an observation video and setting iteration implementation parameters; (2) calculating an iterative optimization solution of a grid point and generating the mixed-state Gauss MRF state value of the grid point according to grid point state judging policy; (3) if the grid point is a moving target point, marking the value of a moving target detection diagram at the grid point and keeping the background reconstruction value of the grid point constant; or else, updating the background and setting the new background reconstruction value of the grid point; (4) generating the state points of all grid points by iteration and generating a mixed-state Gauss MRF state set by utilizing the state points; (5) calculating an overall energy value by utilizing the mixed-state Gauss MRF state set; (6) calculating overall energy change value; if the change value is greater than an iterative error threshold value, continuously performing ICM (Iterated Conditional Modes) iterative optimization; or else, finishing the convergence process of the iterative optimization; and (7) outputting the moving target detection diagram and a background reconstruction diagram. The dynamic updating of the background in the moving target detection process can be realized, the representational capacity of the complex texture background can be improved effectively, the detection precision under the noise environment is improved and the method is applicable to the detection treatment for the moving target of the ink-jet printing fabric.

Description

A kind of ink-jet printed fabric movement object detection method based on mixed state Gauss MRF model

Technical field

The present invention relates to ink-jet printed fabric defect detection field, specifically a kind of moving target detecting method of ink-jet printed fabric.

Background technology

Ink-jet printed fabric a kind ofly adopts the digital ink-jet printed shaping fabric type of high density, it has abandoned the complicated link of Conventional decal plate-making, by controlling the switch of dye nozzles, and by nozzle, dyestuff air brushing on fabric is formed, greatly improve fineness and the rich color degree of fabric picture.Texture levels are rich, the quality exquisiteness of ink-jet printed fabric are true to nature, can show the picture true to nature such as oil painting, landscape painting and grain effect completely, have been widely used in the high-end textile product such as Nanjing brocade, silk.Because ink-jet printed texture dyestuff air brushing on fabric is formed by nozzle, usually have complicated trickle ink institutional framework, the quality of institutional framework is the condition precedent obtaining fine decorative pattern, and it will directly affect the quality of ink-jet printing fabric.

In order to ensure the real-time ink-jet printed drafting effect of nozzle, the fabric knitting video to comprising ink-jet printed defect is needed to monitor.Ink-jet printed fabric defect monitoring is used widely in the industrial detection and quality control of ink-jet printed fabric, becomes one of important method ensureing ink-jet printed fabric product quality.Defect monitoring relates generally to carries out analyzing and processing to the frame sequence in ink-jet printed video, wherein the most key task is defects detection, defects detection is the core technology in video monitoring, and the accuracy of testing result directly will affect subsequent defective monitoring analysis (as Bug Tracking and classification etc.).

Ink-jet printed fabric defect detects can when not needing human intervention, the kinetic characteristic utilizing defect target to show in video scene in time-space domain on difference, be moving target by defect object representation, adopt the moving target detecting method of time-space domain associating that moving target is extracted from video scene, then according to the information such as shape and position of moving target determination defect target, thus automatic detection and the warning of defect abnormal conditions is realized.

The moving target detecting method of current employing time-space domain associating mainly contains three kinds:

(1) frame differential method.Frame differential method by implementing phase reducing to adjacent two two field pictures in sequence image, and extracts motion target area through thresholding process, and the method is implemented simple, and speed is fast, is easy to hardware implementing, and is not very responsive to the change of light.But frame differential method depends on the movement velocity of moving target, excessive velocities or the excessively slow integrity profile being all difficult to acquisition moving target, be easy to produce smear or cavitation.

(2) optical flow method.The sports ground of optical flow method to sequence of frames of video is estimated, merged by similar motion vector and form moving target, but its computation complexity is high, is difficult to realize real-time process.In addition, owing to be subject in real ink-jet printed operating environment the impacts such as noise, irregular and uneven illumination, shade and background perturbation, the method to be completed accurately and the moving object detection work of robust is very difficult.

(3) background null method.The main thought of background null method builds background model, and by comparing incoming frame and background model detects moving target.In recent years, background null method is as the focus in moving target detecting method, and Chinese scholars conducts in-depth research it.Because background null method directly can provide the information such as position, size, shape of moving target, and not by the restriction of target speed, be therefore realize the prefered method that moving target detects in real time and extract.

When the situation that background model is known, background null method is the quite effective moving target detecting method of one.But the method is when facing the change of the background image occurrence dynamics in video scene, and can not reach good effect, Major Difficulties comprises structure and the renewal of background model.In order to solve the difficulty of background null method when background model builds and upgrade, domestic and international many scholars further investigate background model.

At present, the background model adopted in background null method is mainly divided into two classes:

(1) based on the background model of pixel: document [1] D.Lee, " Effective gaussian mixture learning for videobackground subtraction ", IEEE Transactions on Pattern Analysis and Machine Intelligence, vol.27, no.5, pp.827-832, adopt ADAPTIVE MIXED Gauss model to realize time-space domain background modeling, but effect is undesirable in a noisy environment in 2005.Document [2] Zhu Biting, Zheng Shibao. " spatial domain background separation method and shadow removing method based on gauss hybrid models ", Journal of Image and Graphics, 13 (10): 1906-1909, gauss hybrid models is adopted to realize spatial domain background separation and shadow removing in 2008, but very weak to the processing power of time domain scene dynamics change.Document [3] K.K.Hati, P.K.Sa, B.Majhi, " Intensityrange based background subtraction for effective object detection ", IEEE Signal Processing Letters, vol.20, no.8, pp.759-762, adopt the constraints policy arranging background pixel point tonal range in 2013, improve the adaptive faculty that background model changes scene dynamics, but more responsive to noise ratio.

Owing to requiring that each pixel in background is independently based on the background model of pixel, although the profile of moving target effectively can be extracted, but its obtain testing result and scene in the background image degree of correlation very high, when background image occurrence dynamics changes, testing result is just undesirable, simultaneously the rejection ability of the method to noise is more weak, is therefore not suitable in noisy, under dynamically changeable scene ink-jet printed fabric movement target detection.

(2) based on the background model in region: this model can make full use of the mutual relationship in interframe (time domain) and frame in (spatial domain) image between neighbor pixel, therefore the background model based on pixel is better than for the treatment effect of the situations such as noise and dynamically changeable scene.Because Markov random field (Markov Random Field, MRF) can characterize the neighborhood relationships in the same area between neighbor pixel well, therefore many scholars adopt MRF as the background model based on region.Document [4] E.Y.Kim, S.H.Park, " Automatic video segmentation using genetic algorithms ", Pattern RecognitionLetters, vol.27, no.11, pp.1252-1265, utilize the space-time dividing of MRF realization to moving target in 2006, effectively remain the marginal information of target.Document [5] S.S.Huang, L.C.Fu, P.Y.Hsiao, " Region-level motion-basedbackground modeling and subtraction using MRFs ", IEEE Transactions on Image Processing, vol.16, no.5, pp:1446-1456, utilize in 2007 the markov feature of time-space domain improve a gray scale time become ASSOCIATE STATISTICS effect, noise when suppressing motion to detect.Document [6] J.Shen, W.Yang, Z.Lu, et al, " Information integrationfor accurate foreground segmentation in complex scenes ", IET Image Processing, vol.6, no.5, pp.596-605, in 2012, shape priors is introduced MRF background model, effectively improve the accuracy of detection of moving target.

But, larger deficiency is there is: 1) grain background of ink-jet printed fabric is complicated and be full of variety when direct utilization detects ink-jet printed fabric movement target based on the background null method of MRF background model, it is more weak that existing MRF background model characterizes ability to the motion feature of complex texture, and Ability of Resisting Disturbance is poor.2) existing MRF background model is when performing moving object detection, needs to keep background model constant, synchronously cannot perform background reconstruction, lack dynamically updating background model, thus affect the precision of subsequent motion target detection.Because the renewal computing cost of background model when reconstructing is very large, the demand detected in real time cannot be met, be restricted during its moving object detection at ink-jet printed fabric is applied.

Summary of the invention

The present invention is for solving the defect existing for prior art, " motion texture " feature is incorporated in the background modeling of background null method, and set up the mixed state Gauss MRF model that simultaneously can comprise two states type (motion state and background state), a kind of ink-jet printed fabric movement object detection method based on mixed state Gauss MRF model is proposed.The method effectively can not only solve the difficulty of background reconstruction in background null method, realize background dynamics to upgrade, and the accuracy of detection of the sign ability that can improve complex texture background and moving target, thus effectively realize ink-jet printed fabric movement target detection, and there is good robustness to noise.

The technical solution adopted for the present invention to solve the technical problems is:

Based on an ink-jet printed fabric movement object detection method for mixed state Gauss MRF model, comprise the following steps:

Step is 1.: input observation video I, puts time sequence mark t=1, background reconstruction figure u ₀=I ₁, lattice point sequence mark i=0, frame of video sum M=100, lattice point sum N=80 × 80=6400 in single frames, iteration error threshold value η=0.01, iteration sequence mark k=1, global energy value E ₀=0, noise threshold ω=0.1, weighting parameter α=1, β=8, γ=4, variance λ=1.5, smooth scaling factor parameter a=0.7;

Step is 2.: put i=i+1, calculates E (x _t| I, u _t-1) energy value at lattice point i place iteration optimization solution and according to " judgement of lattice point state " strategy, generate the mixed state Gauss MRF state value of lattice point i

Step is 3.: if then marker motion target detection figure m _tbe 1 in the value at lattice point i place, and keep background reconstruction value constant, namely put otherwise renewal background, puts background reconstruction value

Step is 4.: if i<N, go to step 2., continues the state value generating next lattice point; Otherwise, utilize the state value of all lattice points to generate mixed state Gauss MRF state set go to step 5.;

Step is 5.: utilize x _tcalculate E (x _t| I, u _t-1), put global energy value E _k=E (x _t| I, u _t-1), and by energy change value Δ E=E _k-E _k-1as the termination basis for estimation of ICM iteration optimization;

Step is 6.: if Δ E< is η, then the convergence process of ICM iteration optimization completes, and goes to step 7.; Otherwise put k=k+1, i=0, and go to step 2., continue to perform ICM iteration optimization next time;

Step is 7.: utilize optimal estimation value x _t, export the moving object detection figure m of t _twith background reconstruction figure u _t;

Step is 8.: put t=t+1, if t≤M, then put k=1, i=0, go to step 2., continues to perform next frame and detects; Otherwise, complete all frames and detect, terminate.

Described step 1. in, I={I _t| t=1 ..., M} is the two field picture set in all moment, and it is a frame image sequence, wherein, I _t={ I _i(t) | i=1 ..., N} is the two field picture of t, and it is the brightness value set of all lattice points in t, I _it () is the brightness value of lattice point i in t.

Described step 1. in, u={u _t| t=1 ..., M} is the background reconstruction set of graphs in all moment, wherein, be the background reconstruction figure of t, it is the background reconstruction value set of all lattice points in t, the background reconstruction value of lattice point i in t.

Described step 1. in, u ₀u _tvalue when t=0; I ₁i _tvalue when t=1.

Described step 1. in, E ₀global energy value E _kvalue when k=0.

Described step 2. in, E (x _t| I, u _t-1) be the derivation of energy formula of moving object detection model, derivation of energy formula is defined as:

Wherein, x is mixed state stochastic variable, the mixed state Gauss MRF state value of lattice point i in t, be mixed state random field, it is the mixed state Gauss MRF state value set of all lattice points in t, W _ibe the neighborhood point set of lattice point i, j is the neighborhood point set W of lattice point i _iin sequence mark, be the mixed state Gauss MRF state value of neighborhood point j in t of lattice point i, I is observation video, u _t-1the background reconstruction figure in t-1 moment; E _pbeing judgement characteristic energy item, detecting for realizing moving target point; E _qbe reconstruct characteristic energy item, for realizing the brightness value reconstruct of background area point, restructuring procedure will effectively improve the precision of moving object detection; E _rcanonical characteristic energy item, for realizing the regularization smoothing procedure process to motion target area and background area; with e respectively _p, E _qand E _rat the value item at lattice point i place; α, β and γ are the regulating parameter of control 3 bound term effects.

Further, in formula (2) whether be moving target point, namely have if adjudicating lattice point i for the characteristic information relevant according to motion detection:

$E_P^i\left(x_i^t\right|I)=(1-m_i^t)\&CenterDot;V_{\mathrm{MT}}(i,t)---\left(\text{6}\right)$

Wherein, m={m _t| t=1 ..., M} is the moving object detection set of graphs in all moment, be the moving object detection figure of t, it is moving target the showing property value set of all lattice points in t, moving target the showing property value of lattice point i in t; moving target indicative function, this function foundation whether whether be l, making lattice point i is that the property shown of moving target point judges, if that is: then otherwise

V _mT(i, t) is the scalar value of lattice point i " motion texture " feature under t, namely has:

$V_{\mathrm{MT}}(i,t)=V_{\mathrm{avg}}(i,t)\&CenterDot;\frac{{\&dtri;I}_i\left(t\right)}{{\left|\right|{\&dtri;I}_i\left(t\right)\left|\right|}^2}---\left(5\right)$

Wherein, I _it () is the brightness value of lattice point i in t, i _i(t) gradient vector on spatial domain, it is vector modulus value.

Make V (i, t) be the normal direction flow vector of lattice point i in t, namely have

$V(i,t)=-\frac{\frac{{\&PartialD;I}_i\left(t\right)}{\&PartialD;t}\&CenterDot;{\&dtri;I}_i\left(t\right)}{{\left|\right|{\&dtri;I}_i\left(t\right)\left|\right|}^2}---\left(3\right)$

Wherein, i _i(t) gradient vector in time domain.

V in formula (5) _avg(i, t) is lattice point i in the weighted method of t to stream amplitude, namely has

$V_{\mathrm{avg}}(i,t)=\frac{{\mathrm{\&Sigma;}}_{j\&Element;W_i}V(j,t){\left|\right|{\&dtri;I}_j\left(t\right)\left|\right|}^2}{\max ({\mathrm{\&omega;}}^2,{\mathrm{\&Sigma;}}_{j\&Element;W_i}{\left|\right|{\&dtri;I}_j\left(t\right)\left|\right|}^2)}---\left(4\right)$

Wherein, W _ibe the neighborhood point set of lattice point i, j is the neighborhood point set W of lattice point i _iin sequence mark, ω is restraint speckle threshold value, and max is the function getting maximal value in two numbers, and V (j, t) is the normal direction flow vector of neighborhood point j in t of lattice point i, I _jt () is the brightness value of neighborhood point j in t of lattice point i, i _j(t) gradient vector on spatial domain, it is vector modulus value.

Formula (3) is lattice point i " motion texture " feature under t.In order to avoid " motion texture " feature is subject to noise, retain the direction attribute of " motion texture " feature simultaneously, the present invention carries out on average local weighted to formula (3), obtains formula (4), namely lattice point i t weighted method to stream amplitude.In order to the proper vector of simplified style (4) calculates, formula (4) projects along on gradient direction by the present invention, obtains formula (5), i.e. the scalar value of lattice point i " motion texture " feature under t.

Formula comprises a large amount of movable information in (5), can provide the confidence judgement foundation whether lattice point i existing motion conditions.Formula (5) is introduced by the present invention in, and the mixed state Gauss MRF state value of auxiliary judgement lattice point i: when there is motion in lattice point i place, then V _mT(i, t) value will increase severely, and will force the moving target property shown value by energy minimization effect of contraction even the mixed state Gauss MRF state value of lattice point i lattice point i is labeled as moving target point, realizes the detection of moving target point, and by the moving target property the shown value set of all lattice points as the moving object detection figure of t.

Further, in formula (2) for generating the background reconstruction value of lattice point i in t, namely have:

Wherein, j is the neighborhood point set W of lattice point i _iin sequence mark, n is the neighborhood point set W of lattice point i _iin element number; the background reconstruction value of lattice point i in the t-1 moment, the background reconstruction value of neighborhood point j in the t-1 moment of lattice point i, I _jt () is the brightness value of neighborhood point j in t of lattice point i; the moving target indicative function of lattice point i in t, the background indicative function of lattice point i in t; for the background reconstruction value punishment constraint of moving target point, brightness value reconstruct for background area point is estimated; In order to noise decrease is on the impact of background reconstruction value, the present invention will with all impose 3 × 3 window normalization process that variance is λ; for realizing background luminance reconstruction value time domain level and smooth, a is smooth scaling factor regulating parameter.

In formula (7) item only to lattice point i be moving target point (namely ) time effective, and impel background reconstruction value producing enough deviations with the observed reading of n neighborhood point near lattice point i, there is consistent situation with observed reading in punishment background reconstruction value. item only to lattice point i be background dot (namely ) time effective, and impel the state value of this point identical with s functional value.Especially, when initially when cannot provide, a=0 can be made, then

Further, in formula (2) for realizing lattice point i and neighborhood point j ∈ W thereof _ithe regularization smoothing procedure process of region within the jurisdiction, namely has:

Wherein, the moving target indicative function of neighborhood point j in t of lattice point i, the background indicative function of neighborhood point j in t of lattice point i, with be respectively used to the regularization constraint to motion target area and background area, $h\left({\&dtri;I}_i\left(t\right)\right)=\frac{1}{1+{\left|\right|{\&dtri;I}_i\left(t\right)\left|\right|}^2},f(x_i^t,x_j^t)={\left[\frac{255-|x_i^t-x_j^t|}{\mathrm{\&lambda;}}\right]}^2,$ λ is the variance of background area being carried out to 3 × 3 window normalization process, be with between the absolute value of difference.

In formula (8) item only to 2 lattice point i and j be all moving target point (namely ) time effective, h function is for suppressing the brightness generation acute variation at moving target point i place, thus the level and smooth regularization of realization to motion target area. item only to 2 lattice point i and j be all background area point (namely ) time effective, this level and smooth regularization except utilizing h function to realize background area, also utilizes f function to improve the contrast of background area, the generation of the too small situation of deviation of luminance values effectively between adjacent two background dots of punishment.

Described step 2. in, be the derivation of energy formula of lattice point i in t, expression formula is defined as:

$E\left(x_i^t\right|I,u_{t-1})=\mathrm{\&alpha;}{E}_P^i\left(x_i^t\right|I)+\mathrm{\&beta;}{E}_Q^i\left(x_i^t\right|I,u_{t-1})+\mathrm{\&gamma;}{E}_R^{i,j\&Element;W_i}(x_i^t,x_j^t|I)---\left(9\right)$

And all lattice points are the derivation of energy formula (2) of moving object detection model in the derivation of energy formula sum of t, namely have:

$E\left(x_t\right|I,u_{t-1})=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}E\left(x_i^t\right|I,u_{t-1})---\left(10\right)$

Described step 2. in, be adopt the iteration optimization solution that ICM method obtains, iterative solution is defined as:

$x_i^{*}=\frac{\mathrm{\&beta;s}(u_i^{t-1},I_i\left(t\right))+\mathrm{\&gamma;h}\left({\&dtri;I}_i\left(t\right)\right)\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{x}_j^t{m}_l^{*}\left(x_j^t\right)}{\mathrm{\&beta;}+\mathrm{\&gamma;h}\left({\&dtri;I}_i\left(t\right)\right)\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{m}_l^{*}\left(x_j^t\right)}---\left(12\right)$

Described step 3. in, l is-symbol value is moving target point for marking lattice point; be in the value in t-1 moment, namely lattice point i is in the background reconstruction value in t-1 moment; m _tthe value of m in t, i.e. the moving object detection figure of t.

Described step 5. in, E _k-1e _kvalue when k-1 iteration.

Technical conceive of the present invention is:

1) meticulous and irregular for veining structure complex structure in ink-jet printed fabric, form, the simple problem relying on the feature difference such as brightness, shape on conventional time-space domain cannot represent the defect distortion details (particularly ink-jet drop being formed to the nonlinear deformation behavior of complex texture pattern) of local labyrinth, the present invention adopts " motion texture " feature as moving object detection characteristic information, strengthen moving object detection model to the sign ability of complex texture background, effectively moving object detection precision under raising complex texture background;

2) for the deficiency of existing MRF model only status of support value singlet, the present invention adopts the mixed state Gauss MRF model that simultaneously can comprise two states type (motion state and background state), the singlet characteristic of existing MRF model state value is improved, realize the polymorphic sign to the moving target property shown value and background value, effectively solve the difficulty of existing MRF model on the polymorphic characteristic present of process;

3) adopt a kind of moving object detection model based on mixed state Gauss MRF model, support that the background dynamics in detection process of moving target upgrades, solve the difficulty that existing detection method carries out background reconstruction.Moving object detection model is using the prior imformation as model while of video observing measured value and background reconstruction value, and the characteristic energy of Modling model accordingly, implementation model, to the associating energy characterization of moving object detection and background reconstruction process, improves the adaptive faculty of model to noise and background perturbation.Characteristic energy is made up of move judgement characteristic energy item, background reconstruction characteristic energy item and canonical characteristic energy item.Motion judgement characteristic energy item is used for realizing moving target point and detects, background reconstruction characteristic energy item is for realizing the brightness value reconstruct of background area point, restructuring procedure will effectively improve the precision of moving object detection, canonical characteristic energy item, for realizing the regularization smoothing procedure process to motion target area and background area, strengthens the antijamming capability to noise;

4) adopt energy-optimised technology, detection process of moving target is converted into the optimization problem of moving object detection aspect of model energy.The present invention adopts ICM iteration optimization method for solving, introduce " judgement of lattice point state " strategy, successively local optimization operations is carried out to the characteristic energy of each lattice point, by the optimizing iteration to all lattice points, realize the Optimization Solution to moving object detection aspect of model energy, thus complete the detection of moving target.

The present invention adopts " motion texture " feature as moving object detection characteristic information, set up the moving object detection model based on mixed state Gauss MRF model, not only increase the sign ability of detection model to complex texture background, effective raising detection model is to the accuracy of detection at fine textures edge, and achieve the polymorphic sign of detection model to the moving target property shown value and background value, support that the background dynamics in detection process of moving target upgrades, thus can obtain stable, accurate testing result, under being highly suitable for noisy environment, moving object detection is carried out to the ink-jet printed fabric with complex texture structure.

The effect that the present invention is useful is:

1) adopt " weighted method is to stream amplitude " at the projection strategy of gradient direction, build " motion texture " feature with good noise inhibition, and it can be used as moving object detection characteristic information, for moving target judges to provide decision-making foundation, strengthen moving object detection model to the sign ability of complex texture background, moving object detection precision under raising complex texture background.

2) mixed state variable strategy is adopted, foundation can comprise the mixed state Gauss MRF model of two states type (motion state and background state) simultaneously, realize characterizing while the moving target property shown value and background value, support that the background dynamics in detection process of moving target upgrades.

3) characteristic energy based on the moving object detection model of mixed state Gauss MRF model is set up, and be broken down into motion judgement characteristic energy item, background reconstruction characteristic energy item and canonical characteristic energy item, realize the associating energy characterization to moving object detection and background reconstruction process, improve the adaptive faculty of moving object detection model to noise and background perturbation.

4) detection process of moving target is converted into the characteristic energy optimization problem of moving object detection model, adopts ICM iterative optimization method, introduce " judgement of lattice point state " strategy, the Optimization Solution of realization character energy, and the detection completing moving target.In the process performing moving object detection, support that background dynamically updates, synchronously realize background reconstruction, improve the accuracy of detection of moving object detection model under complex texture background and noise circumstance.

Accompanying drawing explanation

Fig. 1 is the flowchart of the inventive method;

Fig. 2 is the monitor video frame sequence (totally 100 frames) extracted from ink-jet printed fabric CAD system, contains the complex texture background of ink-jet printed fabric in the sequence, and existing defects distortion block in background patterns; Wherein Fig. 2 (a)-(f) is the 8th (normally), 16 (hole defects), 35 (oil droplet marking defects), 52 (normally), 79 (scratch defects), 89 (normally) two field picture respectively; Fig. 2 (g)-(i) is defect fact (ground truth) figure of Fig. 2 (b), 2 (c) and 2 (e) respectively, for providing the effectiveness comparison of detection method;

Fig. 3 is that the inventive method carries out the result of moving object detection and background reconstruction to Fig. 2 frame sequence; Fig. 3 (a)-(c) provides moving object detection figure (m corresponding to defective frame in Fig. 2 (i.e. Fig. 2 (b), Fig. 2 (c) and Fig. 2 (e)) respectively ₁₆, m ₃₅and m ₇₉), namely mixed state Gauss MRF state value equals the lattice point set of l, and these points are the motor points be detected, and marked the position that ink-jet drop deformation defect produces; Fig. 3 (d)-(f) shows background reconstruction figure (u corresponding to defective frame in Fig. 2 (i.e. Fig. 2 (b), Fig. 2 (c) and Fig. 2 (e)) ₁₆, u ₃₅and u ₇₉);

Fig. 4 is the noisy image (noise variance σ=20) of Fig. 2.Wherein, Fig. 4 (a), (b), (c), (d) are the noisy image of Fig. 2 (a), (b), (c), (e) respectively;

Fig. 5 is the inventive method and two kinds of typical method for testing motion (the Huang method of the Kim method in document [4] and document [5]) moving object detection results contrast to Fig. 4 (b)-(d) based on MRF background model;

Fig. 6 is under noisy environment (noise variance σ=20), and two width extracted from monitor video comprise frame of video and the defect fact figure thereof of number of drawbacks type simultaneously.Wherein, Fig. 6 (a)-(b) is the frame of video comprising number of drawbacks type, and Fig. 6 (c)-(d) is the defect fact figure of Fig. 6 (a)-(b);

Fig. 7 be the inventive method with based on pixel background model moving target detecting method (the Hati method in document [3]) and introduce the moving target detecting method based on MRF background model (the Shen method of document [6]) the moving object detection results contrast to Fig. 6 (a)-(b) of prior imformation;

Fig. 8 is under different Noise Criterion, and the inventive method and four kinds of different motion object detection methods are to the moving object detection results contrast of defective frame in Fig. 6 (a).

Embodiment

Below in conjunction with accompanying drawing, the invention will be further described.

With reference to Fig. 1 ~ Fig. 8, a kind of ink-jet printed fabric movement object detection method based on mixed state Gauss MRF model, described detection method comprises the following steps:

Step is 1.: input observation video I, puts time sequence mark t=1, background reconstruction figure u ₀=I ₁, lattice point sequence mark i=0, frame of video overall length M=100, lattice point sum N=80 × 80=6400 in single frames, iteration error threshold value η=0.01, iteration sequence mark k=1, global energy value E ₀=0, noise threshold ω=0.1, weighting parameter α=1, β=8, γ=4, variance λ=1.5, smooth scaling factor parameter a=0.7;

Step is 2.: put i=i+1, calculates E (x _t| I, u _t-1) energy value at lattice point i place iteration optimization solution and according to " judgement of lattice point state " strategy, generate the mixed state Gauss MRF state value of lattice point i

Step is 3.: if then marker motion target detection figure m _tbe 1 in the value at lattice point i place, and keep background reconstruction value constant, namely put otherwise renewal background, puts background reconstruction value

Step is 4.: if i<N, go to step 2., continues the state value generating next lattice point; Otherwise, utilize the state value of all lattice points to generate mixed state Gauss MRF state set go to step 5.;

Step is 5.: utilize x _tcalculate E (x _t| I, u _t-1), put global energy value E _k=E (x _t| I, u _t-1), and by energy change value Δ E=E _k-E _k-1as the termination basis for estimation of ICM iteration optimization;

Step is 6.: if Δ E< is η, then the convergence process of ICM iteration optimization completes, and goes to step 7.; Otherwise put k=k+1, i=0, and go to step 2., continue to perform ICM iteration optimization next time;

Step is 7.: utilize optimal estimation value x _t, export the moving object detection figure m of t _twith background reconstruction figure u _t;

Step is 8.: put t=t+1, if t≤M, then put k=1, i=0, go to step 2., continues to perform next frame and detects; Otherwise, complete all frames and detect, terminate.

1. mixed state Gauss MRF model

While realizing motion detection, can implement dynamically update to background area, the singlet characteristic of the present invention to existing MRF model state value is improved, and sets up mixed state Gauss MRF model.Mixed state Gauss MRF model is a kind of MRF model that simultaneously can comprise two states type, a kind of state is the moving target property the shown value (motion state) that record moves change, and another kind of state is the brightness value (background state) of record background area.

Because above-mentioned two class states belong to discrete type and continuous type respectively, and they are not independent distribution in spatial domain and time domain, therefore need to set up admixture (referred to as the mixed state) statistical distribution pattern that simultaneously can comprise discrete type and continuous type.

L is-symbol value is established in definition 1, and L=[0,255] is the successive value representing brightness range, H={l} ∪ L is the mixed state space comprising discrete symbols value and successive value, then make x ∈ H be mixed state stochastic variable, and according to probability ρ ∈ [0,1] make x=l, make x ∈ L according to probability 1-ρ.

Definition 2 establishes x to be mixed state stochastic variable, then make the mixed state stochastic variable of lattice point i under t, and the mixed state stochastic variable set of all lattice points it is the mixed state random field under t.

Lattice point collection S={1 is established in definition 3 ..., N}, the neighborhood point set of lattice point i, and then have if x _tmeet Markov property, namely have: then claim x _tbe mixed state Markov random field, and it can be used as mixed state MRF model.

Definition 4 establishes lattice point i under t, belong to the probability of motion state moving target indicative function $m_l\left(x_i^t\right)=\left\{\begin{array}{c}1,x_i^t=l\\ 0,x_i^t\&NotEqual;l\end{array}\right.;$ Lattice point i belongs to the probability of background state under t background indicative function and meet normal Gaussian distribution σ is Noise Criterion, then claim x _tfor mixed state Gauss MRF model, and the probability density of model

The present invention adopts the mixed state Gauss MRF model in definition 4, and will mix state stochastic variable as the mixed state Gauss MRF state value of lattice point i under t, state random field x will be mixed _tas the mixed state Gauss MRF state value set (referred to as mixed state Gauss MRF state set) of all lattice points, change the singlet characteristic of existing MRF model state value, namely state value both can be the discrete symbols value of characterizing motility order labeled, also can be the successive value characterizing static background brightness, thus solve the difficulty of existing MRF model on the polymorphic characteristic present of process, realize the polymorphic sign to the moving target property shown value and background value.Utilize mixed state Gauss MRF model can realize the synchronous operation of moving object detection and background reconstruction, thus guarantee dynamically updating Gaussian Background model, improve the precision of moving object detection.

2. based on the moving object detection model of mixed state Gauss MRF model

Only rely on video observing measured value for prior imformation in existing moving target detecting method, testing process is subject to the deficiency of noise and change of background impact, and the present invention adopts a kind of moving object detection model based on mixed state Gauss MRF model.Video observing measured value and background reconstruction value simultaneously as the prior imformation of detection model, and are set up the characteristic energy of detection model by moving object detection model accordingly, realize the associating energy characterization of detection model to moving object detection and background reconstruction process.

The characteristic energy of moving object detection model is made up of move judgement characteristic energy item, background reconstruction characteristic energy item and canonical characteristic energy item.Motion judgement characteristic energy item is used for realizing moving target point and detects, background reconstruction characteristic energy item is for realizing the brightness value reconstruct of background area point, restructuring procedure will effectively improve the precision of moving object detection, canonical characteristic energy item, for realizing the regularization smoothing procedure process to motion target area and background area, strengthens the antijamming capability to noise.

The prior imformation of 2.1 moving object detection models

In existing moving target detecting method, prior imformation only comprises video observing measured value usually, because the present invention is while realizing moving object detection, also synchronously will realize background reconstruction, therefore need background reconstruction value also as the prior imformation of moving object detection model.

If be the background luminance reconstruction value of lattice point i in t, the present invention adopts following context update strategy to obtain background reconstruction value: when lattice point i belongs to moving target, by the background luminance reconstruction value in t-1 moment as the background luminance reconstruction value of t, otherwise will as the background luminance reconstruction value of t, namely have:

$u_i^t=\left\{\begin{array}{c}{u}_i^{t-1},x_i^t=l\\ x_i^t,x_i^t\&NotEqual;l\end{array}\right.---\left(1\right)$

The present invention is by the background luminance reconstruction value set of all lattice points in t as the background reconstruction figure of t, by u={u _t| t=1 ..., M} is as the background reconstruction set of graphs in all moment.

The background reconstruction figure in t-1 moment can be obtained by formula (1) using the prior imformation as detection model while of itself and video observing measured value I, then based on mixed state Gauss MRF model x _tthe derivation of energy formula of moving object detection model can be labeled as E (x _t| I, u _t-1).

The characteristic energy item of 2.2 moving object detection models

Because the derivation of energy formula of MRF model is made up of multistage group potential energy collection, in order to meet the demand detected in real time, the present invention simplifies group potential energy collection, the group that only consideration two is adjacent each other, the therefore derivation of energy formula E (x of moving object detection model _t| I, u _t-1) be made up of the group potential energy of second order at the most.Single order group potential energy is arranged to two parts by the present invention: judgement bound term E _pwith Reconstruction Constraints item E _q, second order group potential energy is set to canonical bound term E _r, then E (x _t| I, u _t-1) be expressed as follows:

Wherein, W _ibe the neighborhood point set of lattice point i, j is the neighborhood point set W of lattice point i _iin sequence mark, the mixed state Gauss MRF state value of neighborhood point j in t of lattice point i; E _pbeing judgement characteristic energy item, detecting for realizing moving target point; E _qbe reconstruct characteristic energy item, the brightness value for background area point reconstructs, and restructuring procedure will effectively improve the precision of moving object detection; E _rcanonical characteristic energy item, for carrying out regularization smoothing procedure process to motion target area and background area; with e respectively _p, E _qand E _rat the value item at lattice point i place; α, β and γ are the regulating parameter of control 3 bound term effects.

2.2.1 judgement characteristic energy item

Judgement characteristic energy item whether adjudicate lattice point i for the characteristic information relevant according to motion detection is moving target point.Due to veining structure complex structure in ink-jet printed fabric, form is meticulous and irregular, the simple defect distortion details (particularly ink-jet drop being formed to the nonlinear deformation behavior of complex pattern texture) relying on the feature difference such as brightness, shape on conventional time-space domain cannot represent local labyrinth, therefore must adopt and have the motion detection feature information that higher complex texture characterizes ability.

" motion texture " is a kind of Time-space serial with time-space domain statistical property, and comprising can the moving object detection characteristic information of Efficient Characterization complex texture background.Motion texture is the expansion of static texture in time domain, and static texture only can reflect the variation characteristic on spatial domain, single frames local, and motion texture can reflect the motion feature of frame sequence in overall time domain.Containing the much information relevant with printing quality (as stamp shape and texture etc.) with ink ejection operation (as ink-jet drop size and shape etc.) in motion texture, is the important evidence judging moving target.From a large amount of ink-jet printed sequence of frames of video that reality obtains, the change of motion texture near moving target on position is comparatively obvious, and these motion textures are dispersed in dynamic scene.

The present invention utilizes the time-space domain associating expression behaviour of ink-jet printed sequence of frames of video, adopt " weighted method is to stream amplitude " at the projection strategy of gradient direction, build " motion texture " feature with good noise inhibition, for moving target judges to provide decision-making foundation, and strengthen the sign ability of moving object detection model to complex texture background.

If t ∈ 1 ..., M} is time sequence mark, and M is frame of video sum, i ∈ 1 ..., N} is lattice point sequence mark, and N is lattice point sum in single frames, then I _it () is the brightness value of lattice point i in t, I _t={ I _i(t) | i=1 ..., N} is the two field picture of t, and it is the brightness value set of all lattice points in t, I={I _t| t=1 ..., M} is the two field picture set in all moment, and it is a frame image sequence.Order with i _i(t) gradient vector in time domain and spatial domain, it is vector modulus value, then

$V(i,t)=-\frac{\frac{{\&PartialD;I}_i\left(t\right)}{\&PartialD;t}\&CenterDot;{\&dtri;I}_i\left(t\right)}{{\left|\right|{\&dtri;I}_i\left(t\right)\left|\right|}^2}---\left(3\right)$

Be the normal direction flow vector of lattice point i in t, and it can be used as " motion texture " feature.

In order to avoid " motion texture " feature is subject to noise, the direction attribute of keeping characteristics simultaneously, the present invention carries out on average local weighted to formula (3), and namely lattice point i in the weighted method of t to stream amplitude is

$V_{\mathrm{avg}}(i,t)=\frac{{\mathrm{\&Sigma;}}_{j\&Element;W_i}V(j,t){\left|\right|{\&dtri;I}_j\left(t\right)\left|\right|}^2}{\max ({\mathrm{\&omega;}}^2,{\mathrm{\&Sigma;}}_{j\&Element;W_i}{\left|\right|{\&dtri;I}_j\left(t\right)\left|\right|}^2)}---\left(4\right)$

Wherein, W _ibe the neighborhood point set of lattice point i, j is the neighborhood point set W of lattice point i _iin sequence mark, ω is restraint speckle threshold value, and max is the function getting maximal value in two numbers, and V (j, t) is the normal direction flow vector of neighborhood point j in t of lattice point i, I _jt () is the brightness value of neighborhood point j in t of lattice point i, i _j(t) gradient vector on spatial domain, it is vector modulus value.

In order to the proper vector of simplified style (4) calculates, formula (4) projects along on gradient direction by the present invention, can obtain the scalar value of lattice point i place " motion texture " feature:

$V_{\mathrm{MT}}(i,t)=V_{\mathrm{avg}}(i,t)\&CenterDot;\frac{{\&dtri;I}_i\left(t\right)}{{\left|\right|{\&dtri;I}_i\left(t\right)\left|\right|}^2}---\left(5\right)$

Owing to comprising a large amount of movable information in formula (5), the confidence judgement foundation whether lattice point i existing motion conditions therefore can be provided.The present invention is introduced into in, and the mixed state Gauss MRF state value of auxiliary judgement lattice point i, namely have:

$E_P^i\left(x_i^t\right|I)=(1-m_i^t)\&CenterDot;V_{\mathrm{MT}}(i,t)---\left(\text{6}\right)$

Analyze known to formula (6): when lattice point i place exists motion, then V _mT(i, t) value will increase severely, and will force the moving target property shown value by energy minimization effect of contraction even the mixed state Gauss MRF state value of lattice point i lattice point i is labeled as moving target point, realizes the detection of moving target point.

The present invention is by moving target the showing property value set of all lattice points in t as the moving object detection figure of t, by m={m _t| t=1 ..., M} is as the moving object detection set of graphs in all moment.

2.2.2 reconstruct characteristic energy item

Reconstruct characteristic energy item for generating the background reconstruction value at lattice point i place.From formula (1), lattice point i, by mixed state Gauss MRF state values different for foundation 2 kinds, makes comprise the energy term of 2 not same-actions with for the background reconstruction value punishment constraint of moving target point, brightness value reconstruct for background area point is estimated.In order to noise decrease is on the impact of background reconstruction value, the present invention will with all impose 3 × 3 window normalization process that variance is λ.

Wherein, n is the neighborhood point set W of lattice point i _iin element number, the background reconstruction value of lattice point i in the t-1 moment, the background reconstruction value of neighborhood point j in the t-1 moment of lattice point i, I _jt () is the brightness value of neighborhood point j in t of lattice point i; the moving target indicative function of lattice point i in t, the background indicative function of lattice point i in t; for the background reconstruction value punishment constraint of moving target point, brightness value reconstruct for background area point is estimated; In order to noise decrease is on the impact of background reconstruction value, the present invention will with all impose 3 × 3 window normalization process that variance is λ; for realizing background luminance reconstruction value time domain level and smooth, a is smooth scaling factor regulating parameter.

Analyze known to formula (7): 1. item only to lattice point i be moving target point (namely ) time effective, and impel background reconstruction value producing enough deviations with the observed reading of n neighborhood point near lattice point i, there is consistent situation with observed reading in punishment background reconstruction value.2. item only to lattice point i be background dot (namely ) time effective, and impel the state value of this point identical with s functional value.Especially, when initially when cannot provide, a=0 can be made, then

2.2.3 canonical characteristic energy item

Canonical characteristic energy item for realizing lattice point i and neighborhood point j ∈ W thereof _ithe regularization smoothing procedure process of region within the jurisdiction, it is by 2 energy terms with form, be respectively used to the regularization constraint to motion target area and background area.

Wherein, n is the neighborhood point set W of lattice point i _iin element number, the moving target indicative function of neighborhood point j in t of lattice point i, the background indicative function of neighborhood point j in t of lattice point i, $h\left({\&dtri;I}_i\left(t\right)\right)=\frac{1}{1+{\left|\right|{\&dtri;I}_i\left(t\right)\left|\right|}^2},f(x_i^t,x_j^t)={\left[\frac{255-|x_i^t-x_j^t|}{\mathrm{\&lambda;}}\right]}^2,$ λ is the variance of background area being carried out to 3 × 3 window normalization process, be with between the absolute value of difference.

Analyze known to formula (8): 1. item only to 2 lattice point i and j be all moving target point (namely ) time effective, h function is for suppressing the brightness generation acute variation at moving target point i place, thus the level and smooth regularization of realization to motion target area.2. item only to 2 lattice point i and j be all background area point (namely ) time effective, this level and smooth regularization except utilizing h function to realize background area, also utilizes f function to improve the contrast of background area, the generation of the too small situation of deviation of luminance values effectively between adjacent two background dots of punishment.

3. moving target detecting method

The present invention proposes a kind of moving target detecting method, and detection process of moving target is converted into moving object detection model energy expression formula E (x by the method _t| I, u _t-1) energy-optimised problem, adopt energy-optimised method for solving to realize moving object detection, namely first by minimizing E (x _t| I, u _t-1) obtain x _testimated value, and then generate moving object detection figure m _twith background reconstruction figure u _t, realize the synchronous operation of moving object detection and background reconstruction.

At present, there is a lot of energy-optimised method for solving, document [7] F.Huang, S.Narayan, D.Wilson, et al, " Afast iterated conditional modes algorithm for water-fat decomposition in MRI ", IEEE Transactionson Medical Imaging, vol.30, no.8, pp.1480-1492, ICM (Iterated Conditional Modes) method in 2011 is typical determinacy relaxation method, it is fast that it has computing velocity, the feature restrained can be ensured in less iterations, the Optimization Solution being effective to MRF model calculates.

The present invention adopts ICM iteration optimization method for solving, and introduces " judgement of lattice point state " strategy, carries out local optimization operations successively, by the optimizing iteration to all lattice points, realize E (x the characteristic energy of each lattice point _t| I, u _t-1) Optimization Solution of energy, thus complete the detection of moving target.

The derivation of energy formula (2) of moving object detection model is decomposed, the derivation of energy formula of single lattice point i in t can be obtained:

$E\left(x_i^t\right|I,u_{t-1})=\mathrm{\&alpha;}{E}_P^i\left(x_i^t\right|I)+\mathrm{\&beta;}{E}_Q^i\left(x_i^t\right|I,u_{t-1})+\mathrm{\&gamma;}{E}_R^{i,j\&Element;W_i}(x_i^t,x_j^t|I)---\left(9\right)$

And all lattice points are the derivation of energy formula (2) of moving object detection model in the derivation of energy formula sum of t, namely have:

$E\left(x_t\right|I,u_{t-1})=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}E\left(x_i^t\right|I,u_{t-1})---\left(10\right)$

After carrying out energy-optimised solving to formula (9), lattice point i is in the state of t there are two kinds of value possibilities:

1. when lattice point i is moving target point, state value

2., when lattice point i is background area point, formula (9) will be reduced to:

$E\left(x_i^t\right|I,u_{t-1})=\mathrm{\&beta;}{\left[\frac{x_i^t-s(u_i^{t-1},I_i\left(t\right))}{\mathrm{\&lambda;}}\right]}^2+\mathrm{\&gamma;}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}h\left({\&dtri;I}_i\left(t\right)\right)m_l^{*}\left(x_j^t\right){\left[\frac{255-|x_i^t-x_j^t|}{\mathrm{\&lambda;}}\right]}^2---\left(11\right)$

Formula (11) is made up of several Gauss's items, utilizes Gauss's item average characteristics, adopts the iteration optimization solution of ICM method can be expressed as:

$x_i^{*}=\frac{\mathrm{\&beta;s}(u_i^{t-1},I_i\left(t\right))+\mathrm{\&gamma;h}\left({\&dtri;I}_i\left(t\right)\right)\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{x}_j^t{m}_l^{*}\left(x_j^t\right)}{\mathrm{\&beta;}+\mathrm{\&gamma;h}\left({\&dtri;I}_i\left(t\right)\right)\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{m}_l^{*}\left(x_j^t\right)}---\left(12\right)$

, and it can be used as state value, namely

For two kinds of value possibilities, it is tactful that the present invention introduces " judgements of lattice point state ": keep other lattice point state value to fix, in calculating formula (9) with value, if the former is less than the latter, then put otherwise put namely have:

$x_i^t=\left\{\begin{array}{c}l,E(x_i^t=l)<E(x_i^t=x_i^{*})\\ x_i^{*},E(x_i^t=l)\&GreaterEqual;E(x_i^t=x_i^{*})\end{array}\right.---\left(13\right)$

The meaning of formula (13) is: 1. due to each lattice point state by with in state corresponding to least energy generate, therefore effectively ensure that the characteristic energy of ICM method to each lattice point meets local optimum requirement; 2. realize judging the mixed state Gauss MRF state of each lattice point, for the certification mark of follow-up execution moving target point and the brightness value reconstruct of background area point provide decision-making foundation.This application of policies in ICM method, by the optimizing iteration to all lattice points, effectively can be realized moving object detection by the present invention.

According to above analysis, a kind of ink-jet printed fabric movement object detection method design based on mixed state Gauss MRF model is as follows:

Step is 1.: input observation video I, puts time sequence mark t=1, background reconstruction figure u ₀=I ₁, lattice point sequence mark i=0, frame of video overall length M=100, lattice point sum N=80 × 80=6400 in single frames, iteration error threshold value η=0.01, iteration sequence mark k=1, global energy value E ₀=0, noise threshold ω=0.1, weighting parameter α=1, β=8, γ=4, variance λ=1.5, smooth scaling factor parameter a=0.7;

Step is 2.: put i=i+1, calculates E (x _t| I, u _t-1) energy value at lattice point i place iteration optimization solution and according to " judgement of lattice point state " strategy, generate the mixed state Gauss MRF state value of lattice point i

Step is 3.: if then marker motion target detection figure m _tbe 1 in the value at lattice point i place, and keep background reconstruction value constant, namely put otherwise renewal background, puts background reconstruction value

Step is 4.: if i<N, go to step 2., continues the state value generating next lattice point; Otherwise, utilize the state value of all lattice points to generate mixed state Gauss MRF state set go to step 5.;

Step is 5.: utilize x _tcalculate E (x _t| I, u _t-1), put global energy value E _k=E (x _t| I, u _t-1), and by energy change value Δ E=E _k-E _k-1as the termination basis for estimation of ICM iteration optimization;

Step is 6.: if Δ E< is η, then the convergence process of ICM iteration optimization completes, and goes to step 7.; Otherwise put k=k+1, i=0, and go to step 2., continue to perform ICM iteration optimization next time;

Step is 7.: utilize optimal estimation value x _t, export the moving object detection figure m of t _twith background reconstruction figure u _t;

Step is 8.: put t=t+1, if t≤M, then put k=1, i=0, go to step 2., continues to perform next frame and detects; Otherwise, complete all frames and detect, terminate.

Fig. 1 gives the flowchart of the inventive method;

Fig. 2 gives the monitor video frame sequence extracted from ink-jet printed fabric CAD system (totally 100 frames), contains the complex texture background of ink-jet printed fabric in the sequence, and existing defects distortion block in background patterns; Wherein Fig. 2 (a)-(f) is the 8th (normally), 16 (hole defects), 35 (oil droplet marking defects), 52 (normally), 79 (scratch defects), 89 (normally) two field picture respectively; Fig. 2 (g)-(i) is defect fact (ground truth) figure of Fig. 2 (b), 2 (c) and 2 (e) respectively, for providing the effectiveness comparison of detection method;

Fig. 3 gives the result adopting the inventive method Fig. 2 frame sequence to be carried out to moving object detection and background reconstruction.Fig. 3 (a)-(c) provides moving object detection figure (m corresponding to defective frame in Fig. 2 (i.e. Fig. 2 (b), Fig. 2 (c) and Fig. 2 (e)) respectively ₁₆, m ₃₅and m ₇₉), namely mixed state Gauss MRF state value equals the lattice point set of l, and these points are the motor points be detected, and marked the position that ink-jet drop deformation defect produces.Fig. 3 (d)-(f) shows background reconstruction figure (u corresponding to defective frame in Fig. 2 (i.e. Fig. 2 (b), Fig. 2 (c) and Fig. 2 (e)) ₁₆, u ₃₅and u ₇₉).Testing result (Fig. 3 (a)-(c)) is schemed (Fig. 2 (g)-(i)) with defect fact to be contrasted visible, and the inventive method obtains satisfied Detection results.Because the inventive method adopts context update strategy, luminance reconstruction estimation is carried out to static lattice points all in scene and motion lattice point, avoid background reconstruction figure produces cavity situation at motion lattice point region place.As can be seen from Fig. 3 result, adopt the inventive method can realize the synchronous process of moving object detection and background reconstruction.In addition, because the inventive method adds the confidence judgement foundation of motion conditions in judgement characteristic energy, effectively can suppress the disturbance of dynamic background, improve the Detection results of moving target.

Fig. 4 gives the noisy image (noise variance σ=20) of Fig. 2.Wherein, Fig. 4 (a), (b), (c), (d) are the noisy image of Fig. 2 (a), (b), (c), (e) respectively.

Fig. 5 gives the inventive method and two kinds of typical method for testing motion (the Huang method of the Kim method in document [4] and document [5]) moving object detection results contrast to Fig. 4 (b)-(d) based on MRF background model.Wherein, Fig. 5 (a)-(c) is the testing result of Kim method to Fig. 4 (b)-(d) respectively, Fig. 5 (d)-(f) is the testing result of Huang method to Fig. 4 (b)-(d) respectively, and Fig. 5 (g)-(i) is the testing result of the inventive method to Fig. 4 (b)-(d) respectively.As can be seen from Fig. 5 result, the Detection results of the inventive method is better than other two kinds of methods.In Fig. 5 (a)-(c), because Kim method is estimated the mistake of background area, create intensive false positive and false negative point, therefore cannot obtain correct Detection results.In Fig. 5 (d)-(f), the false negative point number that Huang method produces is less, obtain good Detection results, but there is the defect of excess smoothness, add false positive point number, many background areas are labeled as defect target by mistake, and form connected region with defect target, therefore more weak to the small sized defects target detection capabilities under noisy environment.In Fig. 5 (g)-(i), the inventive method make use of the constraint interactive relation between sign condition and continuous state, by the normalized of reconstruct characteristic energy item, effectively overcome the interference of noise to moving object detection, reduce the number of false negative and false positive point simultaneously.On the one hand, because false negative point number is less, defect area is successfully detected out.On the other hand, owing to decreasing false positive point number, the detection target of more compacting therefore is generated.

Table 1 gives in Fig. 5 and adopts the accuracy of detection value of three kinds of method gained to compare.The inventive method adopts F score (F-Score, FS) index is as the overall judging quota of method for testing motion precision, and this index is precision ratio (Precision Ratio, PR) and recall ratio (Recall Ratio, RR) harmomic mean, namely has: wherein, TP is true positives number (True positives, TP), and FP is false positive number (Falsepositives, FP), and FN is false negative number (False negatives, FN).Situation as can be seen from table: Kim method owing to creating a lot of artifact around defect area, and therefore a lot, PR and RR index is all lower for false positive point and false negative point number.The RR index of Huang method is higher, although cause false positive point number to increase due to excess smoothness, PR index reduces, and FS index is still better than Kim method.Compared with other 2 kinds of methods, because the inventive method achieves the guarantor limit denoising of homogeneous region in background reconstruction process, therefore false positive point is significantly cut down, and PR index is better than Kim method and Huang method.In addition, because the inventive method is when performing background reconstruction, can immediate updating background image, the probability making defect point be mistaken for background reduces, and therefore false negative point is effectively suppressed, and RR index is higher, final FS index is better, achieves and compacts and accurate Detection results.

The accuracy value of table 1 three kinds of methods compares

Fig. 6 gives under noisy environment (noise variance σ=20), and two width extracted from monitor video comprise frame of video and the defect fact figure thereof of number of drawbacks type simultaneously.Wherein, Fig. 6 (a)-(b) is the frame of video comprising number of drawbacks type, and Fig. 6 (c)-(d) is the defect fact figure of Fig. 6 (a)-(b).

Fig. 7 give the inventive method with based on pixel background model moving target detecting method (the Hati method in document [3]) and introduce the moving target detecting method based on MRF background model (the Shen method of document [6]) the moving object detection results contrast to Fig. 6 (a)-(b) of prior imformation.Wherein, Fig. 7 (a) and Fig. 7 (d) is the testing result of Hati method to Fig. 6 (a)-(b) respectively, Fig. 7 (b) and Fig. 7 (e) is the testing result of Shen method to Fig. 6 (a)-(b) respectively, and Fig. 7 (c) and Fig. 7 (f) is the testing result of the inventive method to Fig. 6 (a)-(b) respectively.The testing result of three kinds of methods and defect fact are schemed (Fig. 6 (c)-(d)) and contrasted visible: although Hati method takes implement to pixel tonal range the strategy that retrains, to adapt to the dynamic change of scene, but its Detection results is in a noisy environment undesirable, false positive point number is a lot, and many non-defective pixels are defect area point because of noise pollution by flase drop.Shen method is owing to introducing in background modeling by shape priors, therefore better to defect area (as perforated and the oil droplet region) Detection results of regular shape.But because defect areas many in Fig. 6 (a)-(b) (as scratch region) are irregular, cause prior imformation to lose efficacy, false negative point number increases a lot, and therefore Shen method is not good to the Detection results in these regions.The inventive method is owing to adopting context update strategy, the guarantor limit denoising of homogeneous region is achieved in background reconstruction process, effectively overcome the interference of noise to moving object detection, Detection results is better than Hati method and Shen method, the moving target in defect area can be detected in a noisy environment preferably, and not by the constraint of region shape restriction.

Table 2 gives in Fig. 7 and adopts the detection perform (FS and working time) of three kinds of method gained to compare.FS index shows that the inventive method gained accuracy of detection is high, false positive point and false negative point less, the degree of consistency situation in live Fig. 6 (c)-(d) of the defect area thus detected and defect is better.Because Hati method and Shen method are in order to improve accuracy of detection, need to perform context update operation, and the inventive method synchronously completes background reconstruction process in testing process, need not perform separately context update operation, therefore effectively shortens working time.

The detection perform of table 2 three kinds of methods compares

Fig. 8 gives under different Noise Criterion, the inventive method and four kinds of different motion object detection methods (the Hati method in the Huang method of the Kim method in document [4], document [5], document [3] and the Shen method of document [6]) the moving object detection results contrast to defective frame in Fig. 6 (a).Wherein, Fig. 8 (a) and (g) to be noise variance σ be respectively 40 and 60 noisy figure; Fig. 8 (b) and (h) are that Kim method is to the testing result of Fig. 8 (a) with (g) respectively; Fig. 8 (c) and (i) are that Huang method is to the testing result of Fig. 8 (a) with (g) respectively; Fig. 8 (d) and (j) are that Hati method is to the testing result of Fig. 8 (a) with (g) respectively; Fig. 8 (e) and (k) are that Shen method is to the testing result of Fig. 8 (a) with (g) respectively; Fig. 8 (f) and (l) are that the inventive method is to the testing result of Fig. 8 (a) with (g) respectively.Contrast visible by live to the testing result of Lung biopsy and defect Fig. 6 (c): along with the increase of Noise Criterion, a lot, moving object detection effect is the poorest for the false positive point of Kim method around defect area and false negative point number.Huang method and the Shen method moving object detection effect under large scale noise is better than Kim method.But the false positive point of Huang method is more, there is flase drop situation.The false negative point of Shen method to irregular scratch region place is more, there is undetected situation.Hati method is owing to adopting pixel background modeling, and many non-defective pixels are moving target point because of noise pollution by flase drop, and the Detection results under large scale noise circumstance is undesirable.The inventive method owing to achieving the guarantor limit denoising of homogeneous region in background reconstruction process, and can immediate updating background image, therefore false positive point and false negative point number are effectively suppressed, the moving target detected is more complete and accurate, and the minutia of moving target keeps situation better.

Content described in this instructions embodiment is only enumerating the way of realization of inventive concept; protection scope of the present invention should not be regarded as being only limitted to the concrete form that embodiment is stated, protection scope of the present invention also and conceive the equivalent technologies means that can expect according to the present invention in those skilled in the art.

Claims (5)

1., based on an ink-jet printed fabric movement object detection method for mixed state Gauss MRF model, comprise the following steps:

Step is 1.: input observation video I, puts time sequence mark t=1, background reconstruction figure u ₀=I ₁, lattice point sequence mark i=0, frame of video sum M=100, lattice point sum N=80 × 80=6400 in single frames, iteration error threshold value η=0.01, iteration sequence mark k=1, global energy value E ₀=0, noise threshold ω=0.1, weighting parameter α=1, β=8, γ=4, variance λ=1.5, smooth scaling factor parameter a=0.7;

Step is 2.: put i=i+1, calculates E (x _t| I, u _t-1) energy value at lattice point i place iteration optimization solution and according to " judgement of lattice point state " strategy, generate the mixed state Gauss MRF state value of lattice point i

Step is 3.: if then marker motion target detection figure m _tbe 1 in the value at lattice point i place, and keep background reconstruction value constant, namely put otherwise renewal background, puts background reconstruction value

Step is 4.: if i<N, go to step 2., continues the state value generating next lattice point; Otherwise, utilize the state value of all lattice points to generate mixed state Gauss MRF state set go to step 5.;

Step is 5.: utilize x _tcalculate E (x _t| I, u _t-1), put global energy value E _k=E (x _t| I, u _t-1), and by energy change value Δ E=E _k-E _k-1as the termination basis for estimation of ICM iteration optimization;

Step is 6.: if Δ E< is η, then the convergence process of ICM iteration optimization completes, and goes to step 7.; Otherwise put k=k+1, i=0, and go to step 2., continue to perform ICM iteration optimization next time;

Step is 7.: utilize optimal estimation value x _t, export the moving object detection figure m of t _twith background reconstruction figure u _t;

Step is 8.: put t=t+1, if t≤M, then put k=1, i=0, go to step 2., continues to perform next frame and detects; Otherwise, complete all frames and detect, terminate.

2., as claimed in claim 1 based on the ink-jet printed fabric movement object detection method of mixed state Gauss MRF model, it is characterized in that:

Described step 1. in, I={I _t| t=1 ..., M} is the two field picture set in all moment, and it is a frame image sequence, wherein, I _t={ I _i(t) | i=1 ..., N} is the two field picture of t, and it is the brightness value set of all lattice points in t, I _it () is the brightness value of lattice point i in t;

U={u _t| t=1 ..., M} is the background reconstruction set of graphs in all moment, wherein, being the background reconstruction figure of t, is the background reconstruction value set of all lattice points in t, the background reconstruction value of lattice point i in t;

U ₀u _tvalue when t=0; I ₁i _tvalue when t=1;

E ₀global energy value E _kvalue when k=0;

Described step 2. in, E (x _t| I, u _t-1) be the derivation of energy formula of moving object detection model, derivation of energy formula is defined as:

Wherein, x is mixed state stochastic variable, the mixed state Gauss MRF state value of lattice point i in t, being mixed state random field, is the mixed state Gauss MRF state value set of all lattice points in t, W _ibe the neighborhood point set of lattice point i, j is the neighborhood point set W of lattice point i _iin sequence mark, be the mixed state Gauss MRF state value of neighborhood point j in t of lattice point i, I is observation video, u _t-1the background reconstruction figure in t-1 moment; E _pbeing judgement characteristic energy item, detecting for realizing moving target point; E _qbe reconstruct characteristic energy item, for realizing the brightness value reconstruct of background area point, restructuring procedure will effectively improve the precision of moving object detection; E _rcanonical characteristic energy item, for realizing the regularization smoothing procedure process to motion target area and background area; with e respectively _p, E _qand E _rat the value item at lattice point i place; α, β and γ are the regulating parameter of control 3 bound term effects.

3., as claimed in claim 2 based on the ink-jet printed fabric movement object detection method of mixed state Gauss MRF model, it is characterized in that:

In formula (2) whether be moving target point, namely have if adjudicating lattice point i for the characteristic information relevant according to motion detection:

$E_P^i\left(x_i^t\right|I)=(1-m_i^t)\&CenterDot;V_{\mathrm{MT}}(i,t)---\left(6\right)$

Wherein, m={m _t| t=1 ..., M} is the moving object detection set of graphs in all moment, being the moving object detection figure of t, is moving target the showing property value set of all lattice points in t, moving target the showing property value of lattice point i in t; moving target indicative function, this function foundation whether whether be l, making lattice point i is that the property shown of moving target point judges, if that is: then otherwise

V _mT(i, t) is the scalar value of lattice point i " motion texture " feature under t, namely has:

$V_{\mathrm{MT}}(i,t)=V_{\mathrm{avg}}(i,t)\&CenterDot;\frac{{\&dtri;I}_i\left(t\right)}{{\left|\right|{\&dtri;I}_i\left(t\right)\left|\right|}^2}---\left(5\right)$

Wherein, I _it () is the brightness value of lattice point i in t, i _i(t) gradient vector on spatial domain, it is vector modulus value;

Make V (i, t) be the normal direction flow vector of lattice point i in t, namely have

$V(i,t)=-\frac{\frac{{\&PartialD;I}_i\left(t\right)}{\&PartialD;t}\&CenterDot;{\&dtri;I}_i\left(t\right)}{{\left|\right|{\&dtri;I}_i\left(t\right)\left|\right|}^2}---\left(3\right)$

Wherein, i _i(t) gradient vector in time domain;

V in formula (5) _avg(i, t) is lattice point i in the weighted method of t to stream amplitude, namely has

$V_{\mathrm{avg}}(i,t)=\frac{{\mathrm{\&Sigma;}}_{j\&Element;W_i}V(j,t){\left|\right|{\&dtri;I}_j\left(t\right)\left|\right|}^2}{\max ({\mathrm{\&omega;}}^2,{\mathrm{\&Sigma;}}_{j\&Element;W_i}{\left|\right|{\&dtri;I}_j\left(t\right)\left|\right|}^2)}---\left(4\right)$

Wherein, W _ibe the neighborhood point set of lattice point i, j is the neighborhood point set W of lattice point i _iin sequence mark, ω is restraint speckle threshold value, and max is the function getting maximal value in two numbers, and V (j, t) is the normal direction flow vector of neighborhood point j in t of lattice point i, I _jt () is the brightness value of neighborhood point j in t of lattice point i, i _j(t) gradient vector on spatial domain, it is vector modulus value;

Formula (3) is lattice point i " motion texture " feature under t.In order to avoid " motion texture " feature is subject to noise, retain the direction attribute of " motion texture " feature simultaneously, carry out on average local weighted to formula (3), obtain formula (4), namely lattice point i in the weighted method of t to stream amplitude; In order to the proper vector of simplified style (4) calculates, formula (4) is projected along on gradient direction, obtains formula (5), be i.e. the scalar value of lattice point i " motion texture " feature under t;

Formula comprises a large amount of movable information in (5), can provide the confidence judgement foundation whether lattice point i existing motion conditions; Formula (5) is introduced in, and the mixed state Gauss MRF state value of auxiliary judgement lattice point i: when there is motion in lattice point i place, then V _mT(i, t) value will increase severely, and will force the moving target property shown value by energy minimization effect of contraction even the mixed state Gauss MRF state value of lattice point i lattice point i is labeled as moving target point, realizes the detection of moving target point, and by the moving target property the shown value set of all lattice points as the moving object detection figure of t.

4., as claimed in claim 3 based on the ink-jet printed fabric movement object detection method of mixed state Gauss MRF model, it is characterized in that:

In formula (2) for generating the background reconstruction value of lattice point i in t, namely have:

Wherein, j is the neighborhood point set W of lattice point i _iin sequence mark, n is the neighborhood point set W of lattice point i _iin element number; the background reconstruction value of lattice point i in the t-1 moment, the background reconstruction value of neighborhood point j in the t-1 moment of lattice point i, I _jt () is the brightness value of neighborhood point j in t of lattice point i; the moving target indicative function of lattice point i in t, the background indicative function of lattice point i in t; for the background reconstruction value punishment constraint of moving target point, brightness value reconstruct for background area point is estimated; In order to noise decrease is on the impact of background reconstruction value, the present invention will with all impose 3 × 3 window normalization process that variance is λ; for realizing background luminance reconstruction value time domain level and smooth, a is smooth scaling factor regulating parameter;

In formula (7) item only to lattice point i be moving target point (namely ) time effective, and impel background reconstruction value producing enough deviations with the observed reading of n neighborhood point near lattice point i, there is consistent situation with observed reading in punishment background reconstruction value; item only to lattice point i be background dot (namely ) time effective, and impel the state value of this point identical with s functional value; Especially, when initially when cannot provide, a=0 can be made, then $s(u_i^{t-1},I_i\left(t\right))=I_i\left(t\right).$

5., as claimed in claim 4 based on the ink-jet printed fabric movement object detection method of mixed state Gauss MRF model, it is characterized in that:

In formula (2) for realizing lattice point i and neighborhood point j ∈ W thereof _ithe regularization smoothing procedure process of region within the jurisdiction, namely has:

Wherein, the moving target indicative function of neighborhood point j in t of lattice point i, the background indicative function of neighborhood point j in t of lattice point i, with be respectively used to the regularization constraint to motion target area and background area, $h\left({\&dtri;I}_i\left(t\right)\right)=\frac{1}{1+{\left|\right|{\&dtri;I}_i\left(t\right)\left|\right|}^2},$ $f(x_i^t,x_j^t)={\left[\frac{255-|x_i^t-x_j^t|}{\mathrm{\&lambda;}}\right]}^2,$ λ is the variance of background area being carried out to 3 × 3 window normalization process, be with between the absolute value of difference;

In formula (8) item only to 2 lattice point i and j be all moving target point (namely ) time effective, h function is for suppressing the brightness generation acute variation at moving target point i place, thus the level and smooth regularization of realization to motion target area; item only to 2 lattice point i and j be all background area point (namely ) time effective, this level and smooth regularization except utilizing h function to realize background area, also utilizes f function to improve the contrast of background area, the generation of the too small situation of deviation of luminance values effectively between adjacent two background dots of punishment.

Described step 2. in, be the derivation of energy formula of lattice point i in t, expression formula is defined as:

$E\left(x_i^t\right|I,u_{t-1})=\mathrm{\&alpha;}{E}_P^i\left(x_i^t\right|I)+\mathrm{\&beta;}{E}_Q^i\left(x_i^t\right|I,u_{t-1})+\mathrm{\&gamma;}{E}_R^{i,j\&Element;W_i}(x_i^t,x_j^t|I)---\left(9\right)$

And all lattice points are the derivation of energy formula (2) of moving object detection model in the derivation of energy formula sum of t, namely have:

$E\left(x_t\right|I,u_{t-1})=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}E\left(x_i^t\right|I,u_{t-1})---\left(10\right)$

Described step 2. in, be adopt the iteration optimization solution that ICM method obtains, iterative solution is defined as:

$x_i^{*}=\frac{\mathrm{\&beta;s}(u_i^{t-1},I_i\left(t\right))+\mathrm{\&gamma;h}\left({\&dtri;I}_i\left(t\right)\right)\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{x}_j^t{m}_l^{*}\left(x_j^t\right)}{\mathrm{\&beta;}+\mathrm{\&gamma;h}\left({\&dtri;I}_i\left(t\right)\right)\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{m}_l^{*}\left(x_j^t\right)}---\left(12\right)$

Described step 3. in, l is-symbol value is moving target point for marking lattice point; be in the value in t-1 moment, namely lattice point i is in the background reconstruction value in t-1 moment; m _tthe value of m in t, i.e. the moving object detection figure of t.

Described step 5. in, E _k-1e _kvalue when k-1 iteration.