CN103700080B - Image mismatch based on Skinner probabilistic automaton is to minimizing technology - Google Patents

Abstract

The invention discloses a kind of image mismatch based on Skinner probabilistic automaton to minimizing technology, including: carrying out images match, it is right to obtain mating；Determine the Skinner probabilistic automaton mathematical model of matching result；Randomly drawing coupling from coupling centering according to probability right, and obtain the basis matrix between image, it is right that then this basis matrix bring into all couplings, according to decision function obtain each coupling between pole adjust the distance and mean pole is adjusted the distance；According to weight adjustment function, each coupling to pole adjust the distance and mean pole is adjusted the distance to each coupling to carrying out the adjustment of weight, and calculate each coupling to extraction probability；Judge whether current basal matrix has the most correct number of pairs, and combine three Rule of judgment control algolithm iterative process.Present invention introduces bionics and cognitive psychology, in the case of lacking wrong matching probability priori, propose three kinds of stopping criterion for iteration, improve the autonomy of algorithm.

Description

Image mismatch based on Skinner probabilistic automaton is to minimizing technology

Technical field

The present invention relates to a kind of image mismatch to removing algorithm, be specifically related to a kind of image mismatch based on Skinner probabilistic automaton to minimizing technology (Skinner-Ransac).

Technical background

In the autonomous optical navigation of spacecraft optical alignment, unmanned plane and robot, image matching technology plays the matching effect between central role, and image undoubtedly and directly affects the precision of navigation.But, owing to there is Geometrical change, illumination between image, trembling shake and the impact of the factors such as noise so that the coupling being constantly present partial error in the matching result between image is right, thus has adverse effect on navigation with location.Therefore, how to find error hiding to a step vital during just becoming images match in matching result.At present, mainly there are following three classes for the removal problem of erroneous matching pair between image: linear approach, iterative method and robust method.Linear approach mainly has 7 methods, 8 methods and 8 methods of improvement, and this algorithm is sensitive for Mismatching point, and the perforated also easily causing Mismatching point in the case of error hiding rate is low removes.The arithmetic accuracy of iterative algorithm is higher than linear approach, but poor in timeliness.Robust algorithm is the class that in three class algorithms, effect is best, and achievement based on this class algorithm was more in recent years, such as the M estimation technique, LMeds and Ransac algorithm etc..The M estimation technique is by problem is converted into the Least Square Solution with weighting, remaining difference square is replaced with a remaining difference function, thus suppress big remaining difference for the impact of estimated result, but the method is strong for initial dependence, and for erroneous matching to also sensitive.LMeds is a kind of minimum median algorithm, and this algorithm is by minimizing the transformation model that remaining difference square intermediate value is estimated between image, and it mates having preferable robustness for mistake, but this algorithm is the most applicable when the wrong matching probability in data is more than 50%.Ransac algorithm is respectively provided with obvious advantage in arithmetic accuracy and robustness compared with the first two, and when the probability of coupling wrong in data is more than 50%, the basis matrix between image is it is estimated that come, so this algorithm has been widely used in machine vision.

Ransac algorithm be mainly characterized by coupling to being passively to be chosen by algorithm to carry out the calculating of basis matrix between image, then the basis matrix of estimation is brought into other coupling centering, adjust the distance and make comparisons with the threshold value set by calculating pole it is estimated that this basis matrix is the coincidence rate of these match points, basis matrix the choosing as final mask parameter maximum by choosing coincidence rate after successive ignition.These characteristics makes this algorithm efficiency of algorithm wrong matching rate is big when reduce.Owing to the iterations of algorithm is that the wrong matching rate according to coupling centering is determined in advance, and the wrong matching rate in matching result is unknown as a rule, and this makes again the autonomy of algorithm be deteriorated.2012, Lu Shan, the paper " based on the conforming Fundamental Matrix Estimation of probability sampling " that thunder hero etc. deliver on " controlling and decision-making " proposes a kind of Pre-testing method, for obtaining preferably basis matrix set, and only comprised after iteration correct coupling to sample set, thus calculate the final basis matrix between image.But find through overtesting, when the error probability in matching result is higher, either Pre-testing process, or iterative process all has the highest amount of calculation, and the setting of iterations also to set according to the wrong matching rate of matching result in advance, so preferable effect can not be reached.2009, Liu Kun, in the paper " the stochastical sampling consistency algorithm that probability guides " that Ge Junfeng etc. deliver on " computer-aided design and graphics journal ", propose the utilization basis matrix obtained of every time sampling to rewind and adjust the distance into the corresponding pole of all samples calculating, then utilize inverse that pole adjusts the distance as the method for each sample weighting.Before the method is compared, method improves each sample initiative in extraction process, but do not adjust the distance and carry out specification in the pole to each sample, some errors matrix pole in some error sample is caused to adjust the distance minimum, it is very big that this allows for weighted value, thus cause the result finally obtaining mistake, and the iterations of the method is also to set in advance according to the wrong matching rate in matching result.The patent of invention of Application No. 200810063012.3 proposes a kind of sample weighting method based on sampling results, but the method is also to set in advance for iterations, does not possess autonomy.

In sum, how to make algorithm have a high operational efficiency, and how to make algorithm possess autonomous stopping criterion for iteration under conditions of mistake mates the unknown be two major issues being currently needed for solving.

American Psychological professor Skinner proposed operant conditioning reflex (operantconditioning in 1938, OC) concept, and it is theoretical to propose famous Skinner operant conditioning reflex by pigeon experiment, and scholar later utilizes this theoretical developments to go out probabilistic automaton based on Skinner trained reflex.If being used for improving Ransac algorithm by Skinner probabilistic automaton principle, make each coupling to the same with each animal individual for selecting to have different initiatives according to accumulation feedback before every time, it will to improve the operational efficiency of Ransac algorithm；If the probability rate of change by sample is applied to the stopping criterion for iteration of algorithm, then can improve the autonomy of algorithm.

Summary of the invention

The purpose of the present invention is to propose to a kind of image mismatch based on Skinner probabilistic automaton to removing algorithm (Skinner-Ransac), to each coupling to giving weights, the change of weights is updated along with the feedback of sampling results, so that its sampling probability occurs corresponding change.The present invention utilizes bionics and cognitive psychology that according to current sampling results, match point is carried out corresponding initiative adjustment, thus gradually steps up the sampling probability of correct match point.The present invention proposes three kinds of stopping criterion for iteration in the case of lacking wrong matching probability priori, thus improves the autonomy of algorithm.

A kind of image mismatch based on Skinner probabilistic automaton, to minimizing technology, comprises the following steps:

Step one, carries out the detection of image characteristic point or characteristic area and mates, and it is right to obtain mating.

Step 2, determines the Skinner probabilistic automaton mathematical model Skinner-Ransac of matching result according to images match result.

Ransac-Ransac is nine tuples, and its expression formula is:

Skinner-Ransac={M, W, P, T, O, Inmost, Mo, N, Stc}(1)

M={m_i| i=1,2 ..., n}, m_i={ (x_i,y_i), (x '_i,y′_i) (2)

W={w_i| i=1,2 ..., n}(3)

$P=\left\{p_i\right|i=\mathrm{1,2},\·\·\·,n\},p_i=w_i/\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{w}_j---\left(4\right)$

$T:d_i=\left\{\begin{array}{cc}c,& d_i\&GreaterEqual;c\\ d_i,& d_i<c\end{array}\right.i=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,n---\left(5\right)$

$O\&RightArrow;w_i=\left\{\begin{array}{cc}{w}_i+\min (\mathrm{round}(\stackrel{\&OverBar;}{d}/d_i),H),& \stackrel{\&OverBar;}{d}/d_i>1\\ \max (w_i-\mathrm{1,1})& \stackrel{\&OverBar;}{d}/d_i\&le;1\end{array}\right.i=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,n---\left(6\right)$

Wherein, M is to mate between image set, m_iIt is right to mate for the i-th between image, and including the corresponding point coordinate of two images, n is a total number of pairs；W be each coupling to weights, initial value w_i=1；P be each coupling to extraction probability, p_iProbability be 1/n；T be judge each coupling to be whether correct coupling under current basal matrix to decision function, and to all erroneous matching between pole adjust the distance d_i(i=1,2 ..., n) giving a constant c, this constant typically can rule of thumb be determined before algorithm performs；O is for the coupling adjustment function to weight, according toValue to each coupling to weights be adjusted,Adjust the distance for mean pole, equal to d_iMeansigma methods, round is the function that rounds up, and Η is maximum award degree；Inmost is used to record and currently at most correctly mates quantity, and initial value is 0；Mo is used for storing currently at most correctly mating the basis matrix corresponding to logarithm, and initial value is 3 × 3 full null matrix；N is algorithm maximum iteration time, when iterations is more than N, will stop iteration, and generally N will be set to 1000；Stc is the iteration stopping condition in probabilistic automaton, utilizes the coupling situation of change end condition as algorithm to extracting probability in the present invention.

Step 3, m coupling is randomly drawed to (m is 7 or 8) according to probability from coupling centering, and obtaining the basis matrix between image according to m point method, it is right that then this basis matrix bring into all couplings, according to the decision function T in probabilistic automaton obtain each coupling between pole adjust the distance d_iAdjust the distance with mean pole

Step 4, according to the weight adjustment function O in probabilistic automaton, each coupling to pole adjust the distance d_iAdjust the distance with mean poleTo each coupling to carrying out the adjustment of weight, and according to each coupling of weight calculation after adjusting to extraction probability P.

Step 5, it is judged that whether current basal matrix has the most correct number of pairs, if it did not, go to step three；If it has, and meet one of following condition and just terminate algorithm iteration:

Condition 1:

$K\&GreaterEqual;\frac{\log (1-p)}{\log (1-{(\mathrm{Inmost}/n)}^m)}---\left(7\right)$

Wherein, p is confidence probability, m be every time extraction coupling to number.

Condition 2: iteration stopping condition Stc in probabilistic automaton, to mate the end condition as algorithm of the situation of change to extraction probability, its expression formula is:

$\frac{\underset{j=K-L+1}{\overset{K}{\mathrm{\&Sigma;}}}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}\left|(p_i^j-p_i^{j-1})\right|}{L}\&le;\mathrm{\&lambda;}---\left(8\right)$

Wherein, K is current iteration number of times；L is step-length, is typically set at 10；λ is probability change threshold, is a positive number the least, typically sets λ as 0.01.

Condition 3: current iteration number of times K is equal to automat maximum iteration time N, and its expression formula is:

K=N(9)

If conditions above is unsatisfactory for entirely, then go to step three.

In step one, images match can select different feature point extraction and matching algorithm according to practical situation, and the coupling that its object is to form precision as far as possible high is right, reduces erroneous matching pair, improves the removal accuracy rate of error hiding pair.

M point method in step 3 is for the visual angle change in three-dimensional environment, owing to the coordinate at image midpoint is two dimension, need to each coupling to two point coordinates give the coordinate of Z-directions, usually constant 1.

Coupling in step 4 to weight Tuning function according to each coupling to the adaptedness of current basal matrix being adjusted the coupling size to weight, the size of adaptedness byBeing given, what adaptedness was good gives to reward accordingly, determines award size according to adaptedness；For adaptedness bad give punish accordingly, punishment degree equal.

In step 5, the confidence probability p in first end condition to run to advance to go at algorithm and to be manually set, and typically takes p=0.98.

Compared with prior art, it is an advantage of the current invention that:

(1) during the thought of Skinner probabilistic automaton is applied to the coupling elimination of image mistake by the present invention, add psychology, bionic thought, each sample feedback for sampling results each time as same animal individual is made to make the adjustment in corresponding initiative, the sample of sample correct after iteration for several times and mistake creates significantly differentiation in initiative, thus improves the operational efficiency of algorithm.

(2) present invention proposes three kinds of stopping criterion for iteration to correctly mating to probability unknown problem, thus improves the autonomy of algorithm.

Accompanying drawing explanation

Fig. 1 is method flow diagram involved in the present invention；

Fig. 2 is the present invention required mean iterative number of time curve chart under different wrong matching rates；

Fig. 3 is Ransac algorithm required mean iterative number of time curve chart under different wrong matching rates；

Fig. 4 is the present invention and Ransac algorithm averagely correctly mating recall ratio comparison diagram under Bu Tong wrong matching rate；

Fig. 5 be the present invention correctly mate to and the relation curve of the erroneous matching sampling probability to drawing and iterations；

Fig. 6 is the two width true pictures shot Same Scene with different view；

Fig. 7 is to use the Sift algorithm matching effect figure to Fig. 6；

Fig. 8 is correct the number of pairs comparison diagram that the present invention and Ransac algorithm are drawn under different iterationses；

Fig. 9 is correctly mating extraction effect after using the method for the invention iteration 300 times.

Detailed description of the invention

The present invention will be further described with detailed description of the invention below in conjunction with the accompanying drawings.

The flow chart of the method for the invention is as it is shown in figure 1, mainly include following step:

Step one, carries out the detection of image characteristic point or characteristic area and mates, and it is right to obtain mating.

Images match often selects suitable feature point extraction and matching algorithm according to the practical situation of image, its object is to can effectively complete feature point extraction under various circumstances and mate, before error hiding is to removing, it is possible to ensure that images match result has high correct matching rate.For the image that shooting interval is shorter, owing to difference is little between image, Harris angle point grid and matching algorithm that real-time is prominent can be selected.For the image that difference between shooting interval length and image is big, Sift, Surf, PCA-Surf and Mser scheduling algorithm good for affine transformation robustness can be selected.

Step 2, determines the Skinner probabilistic automaton mathematical model Ransac-Ransac of matching result according to images match result.

Ransac-Ransac is nine tuples, its expression formula and the implication such as formula (1)～(7) of each amount.

Step 3,7 or 8 couplings are randomly drawed according to probability right from coupling centering, and obtaining the basis matrix between image according to 7 methods or 8 methods, it is right that then this basis matrix bring into all couplings, according to the decision function T in probabilistic automaton obtain each coupling between pole adjust the distance d_iAdjust the distance with mean poleThe visual angle change that 7 methods and 8 methods are directed in three-dimensional environment, owing to the coordinate at image midpoint is two dimension, need to each coupling to two point coordinates give the coordinate of Z-directions, usually constant 1.

Step 4, according to the weight adjustment function O in probabilistic automaton and each coupling to pole adjust the distance d_iAdjust the distance with mean poleTo often mating the adjustment carrying out weight, and according to each coupling of weight calculation after adjusting to extraction probability P.Mate weight Tuning function according to each coupling the adaptedness of current basal matrix being adjusted the coupling size to weight, the size of adaptedness byBeing given, what adaptedness was good gives to reward accordingly, determines award size according to adaptedness；For adaptedness bad give punish accordingly, punishment degree equal.

Step 5, it is judged that whether current basal matrix has the most correct number of pairs, without then turning to step 2, if meeting one of formula (7)～(9) three conditions just terminate algorithm iteration.If three conditions are unsatisfactory for full, then algorithm branches step 2.

Two application examples of the present invention are given below.

Experiment 1: analog data is tested

Analog data is that one group of coupling meeting certain affine transformation is right, to each coupling to applying the Gaussian noise that variance is three pixels, number of pairs is 1000, wrong coupling is respectively set to 800,700,600,500,400,300 to quantity, and i.e. wrong matching rate is respectively 80%, 70%, 60%, 50%, 40% and 30%.In experimentation, 8 methods are utilized to calculate basis matrix, the Skinner-Ransac model proposed according to the present invention, confidence probability p is set to 0.98, maximum iteration time is 1000, each coupling to initial weight be 1, probability is 1/1000, and the pole of the wrong coupling c that adjusts the distance is set to 1.5 pixels, each coupling to maximum range of reward Η be 30, punishment amplitude is steady state value 1, and step-length L in end condition Stc is set to 10, and threshold value λ is set to 0.01.The Skinner-Ransac algorithm and the Ransac algorithm that use the present invention carry out 20 tests respectively to different error rate, calculate two kinds of algorithms at the average time of different error rate and averagely correctly to mate recall ratio, Fig. 2 gives the mean iterative number of time result of Skinner-Ransac algorithm of the present invention, Fig. 3 gives the mean iterative number of time result of Ransac algorithm, Fig. 4 gives and averagely correctly mates recall ratio comparison diagram, Fig. 5 gives Skinner-Ransac algorithm of the present invention and correctly mates when data mistake matching rate is 80% the trendgram increased along with iterations with the erroneous matching sampling probability to drawing.Test result indicate that, the real-time of the present invention and accuracy are superior to Ransac algorithm, and still have preferable effect in the case of data error rate is big.

Experiment 2: true picture is tested

True picture is selected from the two width images shot Same Scene with different view, as shown in Figure 6.The extraction of characteristic point utilizes Sift algorithm with mating, the coupling image that obtains as it is shown in fig. 7, this matching result to have 1340 right.In experimentation, 8 methods are utilized to calculate basis matrix, the Skinner-Ransac model proposed according to the present invention, confidence probability p is set to 0.98, maximum iteration time is 1000, each coupling to initial weight be 1, the most each coupling to probability be 1/n, n is a number of pairs, the pole of the wrong coupling c that adjusts the distance is set to 1.5 pixels, each coupling to maximum range of reward Η be 30, punishment amplitude is steady state value 1, in end condition Stc, long L is set to 10, and threshold value λ is set to 0.01.Fig. 8 gives correct the number of pairs comparison diagram that the Skinner-Ransac algorithm of the present invention and Ransac algorithm are drawn under different iterationses.Every kind of algorithm carries out 20 tests under different iterationses, then takes its meansigma methods, and Fig. 9 gives correctly mating extraction effect after utilizing Skinner-Ransac algorithm iteration 300 times.Test result indicate that, the present invention has higher accuracy rate compared with Ransac algorithm under limited iterations.

Claims (6)

1. an image mismatch based on Skinner probabilistic automaton is to minimizing technology, it is characterised in that comprise the following steps:

Step one, carries out the detection of image characteristic point or characteristic area and mates, and it is right to obtain mating；

Step 2, determines the Skinner probabilistic automaton mathematical model of matching result according to images match result, and its expression formula is:

Skinner-Ransac={M, W, P, T, O, Inmost, Mo, N, Stc}

M={m_i| i=1,2 ..., n}, m_i={ (x_i,y_i),(x′_i,y′_i)}

W={w_i| i=1,2 ..., n}

$P=\left\{p_i\right|i=1,2,...,n\},p_i=w_i/\underset{j=1}{\overset{n}{\&Sigma;}}{w}_j$

$T:d_i=\left\{\begin{array}{cc}c,& d_i\&GreaterEqual;c\\ d_i,& d_i<c\end{array}\right.,i=1,2,...,n$

$O\&RightArrow;w_i=\left\{\begin{array}{cc}{w}_i+min\left(round\right(\stackrel{\&OverBar;}{d}/d_i),H),& \stackrel{\&OverBar;}{d}/d_i>1\\ \max (w_i-1,1)& \stackrel{\&OverBar;}{d}/d_i\&le;1\end{array}\right.,i=1,2,...,n$

Wherein, M is to mate between image set, m_iIt is right to mate for the i-th between image, and including the corresponding point coordinate of two images, n is a total number of pairs；W be each coupling to weights, initial value w_i=1；P be each coupling to extraction probability, p_iProbability be 1/n；T be judge each coupling to be whether correct coupling under current basal matrix to decision function, and to all erroneous matching between pole adjust the distance d_iGiving a constant c, this constant is determined before algorithm performs；O is for the coupling adjustment function to weight, according toValue to each coupling to weights be adjusted,Adjust the distance for mean pole, equal to d_iMeansigma methods, round is the function that rounds up, and H is maximum award degree；Inmost is used to record and currently at most correctly mates quantity, and initial value is 0；Mo is used for storing currently at most correctly mating the basis matrix corresponding to logarithm, and initial value is 3 × 3 full null matrix；N is algorithm maximum iteration time, when iterations is more than N, will stop iteration, N is set to 1000；Stc is the iteration stopping condition in probabilistic automaton, utilizes the coupling situation of change end condition as algorithm to extracting probability；

Step 3, m coupling is randomly drawed according to probability right from coupling centering, and obtaining the basis matrix between image according to m point method, it is right that then this basis matrix bring into all couplings, according to the decision function T in probabilistic automaton obtain each coupling between pole adjust the distance d_iAdjust the distance with mean pole

Step 4, according to the weight adjustment function O in probabilistic automaton, each coupling to pole adjust the distance d_iAdjust the distance with mean poleTo each coupling to carrying out the adjustment of weight, and according to each coupling of weight calculation after adjusting to extraction probability P；

Step 5, it is judged that whether current basal matrix has the most correct number of pairs, if it did not, go to step three；If it has, and meet one of following condition and just terminate algorithm iteration:

Condition 1:

$K\&GreaterEqual;\frac{log(1-p)}{log(1-{\left(Inmost/n\right)}^m)}$

Wherein, p is confidence probability, m be every time extraction coupling to number；

Condition 2: iteration stopping condition Stc in probabilistic automaton, to mate the end condition as algorithm of the situation of change to extraction probability, its expression formula is:

$\frac{\underset{j=K-L+1}{\overset{K}{\&Sigma;}}\underset{i=1}{\overset{n}{\&Sigma;}}\left|(p_i^j-p_i^{j-1})\right|}{L}\&le;\mathrm{\&lambda;}$

Wherein, K is current iteration number of times；L be step size settings be 10；λ is probability change threshold, is a positive number, sets λ as 0.01；

Condition 3: current iteration number of times K is equal to automat maximum iteration time N, and its expression formula is:

K=N

If conditions above is unsatisfactory for entirely, then go to step three.

A kind of image mismatch based on Skinner probabilistic automaton the most according to claim 1 is to minimizing technology, it is characterized in that, in described step one, images match can select different feature point extraction and matching algorithm, the coupling that its object is to form precision as far as possible high is right, reduce erroneous matching pair, improve the removal accuracy rate of error hiding pair.

A kind of image mismatch based on Skinner probabilistic automaton the most according to claim 1 is to minimizing technology, it is characterised in that the m in described step 3 is equal to 7 or 8.

A kind of image mismatch based on Skinner probabilistic automaton the most according to claim 1 is to minimizing technology, it is characterized in that, m point method in described step 3 is for the visual angle change in three-dimensional environment, owing to the coordinate at image midpoint is two dimension, need to each coupling to two point coordinates give the coordinate of Z-directions, it is constant 1.

A kind of image mismatch based on Skinner probabilistic automaton the most according to claim 1 is to minimizing technology, it is characterized in that, coupling in described step 4 to weight Tuning function according to each coupling to the adaptedness of current basal matrix being adjusted the coupling size to weight, the size of adaptedness byBeing given, what adaptedness was good gives to reward accordingly, determines award size according to adaptedness；For adaptedness bad give punish accordingly, punishment degree equal.

A kind of image mismatch based on Skinner probabilistic automaton the most according to claim 1 is to minimizing technology, it is characterised in that in described step 5, the confidence probability p in first end condition to run to advance to go at algorithm and to be manually set, and takes p=0.98.