CN101504531A - Multi-sample parallel estimation method for improving matching precision of holographic correlator - Google Patents

Abstract

The invention relates to a multi-sample parallel estimating method for improving the matching precision of a volume holographic correlator, which belongs to the field of optical information processing. The method comprises the following steps: storing image 1 to image N and image 1' to image N' of library images in the volume holographic correlator respectively; inputting a real-time image S into the volume holographic correlator; determining an estimation variables of x1 to xk, and determining the sampling related point number g; according to a first estimating rule, providing estimating functions and an estimating equation related with the related points; according to a second estimating rule, determining a reference estimation image, and providing the relation between each estimation function; according to a third estimating rule, providing a complete estimating equation, and obtaining the state difference of the real-time image relative to the library images according to the estimating equation; and finally, obtaining the physical state of the real-time image in a k dimension state. The method can improve the identification accuracy under the condition that the number of the library images is definite; and if the identification accuracy is kept unchanged, the number of the library images required to be stored can be greatly reduced. Simultaneously, the method can improve the accuracy instantaneously, and greatly improve the efficiency.

Description

Improve the multi-sample parallel estimation method of matching precision of holographic correlator

Technical field

The present invention relates to a kind of multi-sample parallel estimation technique that body holography correlator improves precision fast that is used for, can be used for images match, Target Recognition etc., belong to the optical information processing field.

Background technology

The holographic correlation technique of body derives from high density body holographic storage technology and optical correlation technology, as shown in Figure 1, body holography correlator is by input picture 100, fourier transform lens 101, body holographic storage material 102, inverse-Fourier transform lens 103, the output face 104 of body holography correlator and the storehouse image 105 that is stored in 102 are formed.The holographic correlation technique of body is stored the Fourier transform spectrum of multiple image 105 in having certain thickness body holographic storage material 102 by multiplex technique, constitute a multiple filter, and replaced spatial filter plate in the Vander Lugt optical correlators with this multiple filter, body holography correlator just can carry out computing cross-correlation (multi-channel parallel is relevant) with multiple image (storehouse image) in being stored in material simultaneously with the image 100 of input like this.Theoretical analysis shows that the optical field distribution in body holography correlator output face 104 is

${\mathrm{<figure-callout\; id="0"\; label="dx"\; filenames="A200810227572E00152.png,A200810227572E00171.png"\; state="\{\{state\}\}">dx</figure-callout>}}_0{\mathrm{dy}}_{0}f\&prime;(x_0,y_0)f_m^{*}(x_0+\mathrm{\&xi;},{\mathrm{<figure-callout\; id="0"\; label="y"\; filenames="A200810227572E00152.png,A200810227572E00171.png"\; state="\{\{state\}\}">y</figure-callout>}}_0+\mathrm{\&eta;})\&times;$

$g(x_c,y_c)\&Proportional;\underset{m=-M}{\overset{M}{\mathrm{\&Sigma;}}}{\&Integral;}_{t\sin c}\left\{\frac{t}{2\mathrm{\&pi;}}\right[k_{\mathrm{mz}}-k_{\mathrm{dz}}+\frac{\mathrm{\&pi;}}{\mathrm{\&lambda;}}\frac{\mathrm{\&xi;}(2x_0+\mathrm{\&xi;})+\mathrm{\&eta;}({2\mathrm{<figure-callout\; id="0"\; label="y"\; filenames="A200810227572E00152.png,A200810227572E00171.png"\; state="\{\{state\}\}">y</figure-callout>}}_0+\mathrm{\&eta;})}{f^2}\left]\right\}$

As can be seen from the above equation, because the existence of material thickness t makes the cross correlation function of input picture and every width of cloth storehouse image all be subjected to the modulation of sinc function, that is " secondary lobe " in the simple crosscorrelation distribution (except that true origin have a few) is suppressed.That is to say each single passage is compared with the traditional optical correlator that the relevant peaks of the cross correlation function of output has narrowed down greatly, this just makes that under certain channel spacing can walk abreast in the output face does not have the correlated results of each passage of output with crosstalking.

The existence of secondary lobe makes channel spacing be restricted, and this has influenced the parallel channel number of body holography correlator greatly, also can influence the output accuracy of each passage.By increasing methods such as material thickness, out of focus storage, random phase modulation, suppressed sidelobes further, the relevant peaks of sharpening related function, thus improve parallel channel density and output accuracy significantly.When secondary lobe is pressed abundantly, can think that the output of body holography correlator just is approximately well

$g(x_c,y_c)\&Proportional;\underset{m=-M}{\overset{M}{\mathrm{\&Sigma;}}}\&Integral;{\mathrm{<figure-callout\; id="0"\; label="dx"\; filenames="A200810227572E00152.png,A200810227572E00171.png"\; state="\{\{state\}\}">dx</figure-callout>}}_0{\mathrm{dy}}_{0}f\&prime;(x_0,y_0)f_m^{*}\left(x_0{,y}_0\right)$

At this moment, the output of correlator deteriorates to a dot matrix, and the brightness of each luminous point is proportional to the inner product value of input picture and storehouse image.Because do not had secondary lobe to crosstalk between the reference point, channel spacing can reduce significantly, the parallel channel number increases greatly, and this makes body holography correlator become the parallel inner product operation device of a kind of optical multichannel.This multi-channel parallel relevant feature of body holography correlator makes it all obtain using widely in fields such as multiple goal identification, images match, optical neural network, database retrieval.

Often demand side is to such class situation in the application of body holography correlator, and each opens certain state of corresponding one or several physical quantity of figure in real time, for example takes translation, rotation, the convergent-divergent of a moving object in real time, or the scene of a bomb blast or the like.Traditional body holography correlator when identification mainly based on " hitting-miss " principle, promptly find out one or several the brightest reference point, think real-time input image and this one or several to put pairing storehouse image the most similar.Based on such principle, in theory, want accurately to determine a pairing physical state of figure in real time, need with the situation that might occur all deposit in the correlator as the storehouse image.But the parallel computation port number of body holography correlator can't hold all storehouse images generally in several thousand magnitudes at present.Can be by original storehouse image be extracted, the image that promptly only takes out the representative state of some correspondences deposits correlator in and solves this contradiction.But so, if still based on traditional " hitting-miss " principle, can only be with " location " accuracy limitations of real-time figure on the precision of storehouse image segmentation.Illustrate, for a parallel mobile object on two dimensional surface, if only image A and its image after ten pixels of right translation (note image herein is an image B) and downward ten pixels of translation (note image herein is an image C) with a certain coordinate position of correspondence deposits correlator in, Shi Bie result just can only have following four kinds of situations so: first kind of situation is A, second kind of situation is B, the third situation is C, the 4th kind of situation be three all non-.This moment, bearing accuracy was exactly ten pixels.Though and utilize interpolation technique can improve the body holography correlator accuracy of identification, but it only uses two points or three points of minority, it is limited to improve precision.By adopting the multi-sample parallel estimation technique of introducing in this patent, not only the coordinate of those the real-time figure between A, B, C can be further determined, and certain interference can be resisted, improved the image recognition precision of body holography correlator greatly.

The signal estimation theory is to study in the noise background, by the observation to signal, how to construct the best estimator problem for the treatment of estimated parameter.In the signal estimation theory, be divided into Bayesian Estimation, maximal possibility estimation, Minimum Mean Square Error estimation, least-squares estimation etc.Multisample is estimated, promptly estimates by same detection source being carried out the multidetector hyperchannel, can effectively improve the precision that detection source is surveyed, for example according to the repeated detection theorem:

If m independent observation of received signal is r ₁r ₂... r _m, each noise sample n _i, i=1,2 ... m is independent identically distributed N (A, a σ n) gaussian variable, and the two statistics of noise sample and useful signal sample is independent.

Have so: $p\left(r_i\right)=\frac{1}{\sqrt{2\mathrm{\&pi;}}{\mathrm{\&sigma;}}_n}\exp (-\frac{{(r_i-A)}^2}{2{\mathrm{\&sigma;}}_n^2})$ The single likelihood function

$p\left(\hat{r}\right)={\left(\frac{m}{2\mathrm{\&pi;}{\mathrm{\&sigma;}}_n^2}\right)}^{1/2}\exp (-\frac{m{(r_i-A)}^2}{2{\mathrm{\&sigma;}}_n^2})$ Likelihood function repeatedly

Likelihood function becomes sharply with the number of times of test, and its rate of change is

So also just can estimate to improve the precision of surveying by signal.

The statistical estimate theory of signal all can be used in many disposal systems that are used for information extraction.These systems comprise radar system, communication system, voice signal processing, Flame Image Process, biomedicine, control automatically, seismology etc.For example, in Flame Image Process, just utilize multisample to estimate to remove noise.

The multisample estimation technique is applied to body holography correlator, can well utilizes the characteristics of body holography correlator high-speed parallel, the brightness of the relevant reference point that obtains with the storehouse image of real-time input image as the sample node, is realized that multi-sample parallel estimates.The purpose that multi-sample parallel is estimated is the state for pairing image of this brightness value of more accurate judgement or data page, thereby can improve the precision of the holographic identification of body to a great extent.On the one hand, under the certain situation of storehouse amount of images, can improve accuracy of identification; On the other hand,, then can significantly reduce the storehouse amount of images of required storage, reserve more storage spaces to body holography correlator and hold more diversified storehouse image if keep accuracy of identification constant.Simultaneously, the characteristics of the holographic relevant identification high-speed parallel of body can make the raising moment realization of precision, have improved efficient greatly.

Summary of the invention

As mentioned above, when handling the data of high speed, magnanimity, there is the not high enough problem of precision owing to be subjected to the restriction of its intrinsic operational pattern in body holography correlator, and this precision is not high often because the existence of disturbing.The present invention reaches the anti-interference and comparatively accurate purpose of judging relative status between real-time figure and storehouse image by to carrying out the way of certain multi-sample parallel operation between the storehouse image that is input to the realtime graphic in the correlator and has stored.Therefore, under the certain situation of storehouse amount of images, the present invention can effectively improve the output accuracy of correlator, and is same, and under the situation of equal accuracy of identification, the method for using the present invention to propose can significantly reduce the storehouse picture number that needs storage.Simultaneously, this method can improve the correlator anti-jamming capacity efficiently.

The present invention is adapted to the situation of k dimension (k physical quantity variation, such as translation, rotation, convergent-divergent etc.).

If the storehouse image of having stored is the N width of cloth, import 1 width of cloth realtime graphic, the holographic correlation technique of this input picture and storehouse imagery exploitation body is carried out related operation.Definition N width of cloth storehouse image is respectively Fig. 1 ... figure N schemes S in real time, and the breadth of every width of cloth figure is P * Q, and (unit: pixel count), wherein the state of S is determined by k physical quantity (translation, rotation, convergent-divergent etc.).After figure carries out related calculation with the storehouse image in real time, obtain a series of bright reference points that secretly do not wait, in this example, the reference point number is N.The luminance characterization of reference point the result of related operation.Relatively this result just can obtain the relative status between input figure and storehouse image.Traditional method is the directly relatively brightness of these reference points, and the state of getting the brightest point is approximately virtual condition, and the state that is about to the brightest pairing storehouse of reference point image is defaulted as the state of input picture.As seen, the precision of this method is limited by the gap size of institute's warehousing image, so, if will under certain storehouse image stored number, further improve precision, must seek other method.According to mentioned above, in order to obtain higher precision, the present invention utilizes the multi-sample parallel estimation technique, figure state during with k argument table levies in kind, make full use of the numerical value of reference point brightness, determine k variable-value, thereby obtain the comparatively accurate state of input picture according to estimate equation.

Determine to estimate rule one earlier, estimate rule two, estimate rule three before multisample estimates operation carrying out.

Concrete rule is as follows:

Estimate rule one:

The problem that above-mentioned k dimension continuous state changes (k variable) can be summed up as following problem: for the N width of cloth storehouse image that is stored in the body holography correlator, figure S is in real time arranged arbitrarily, the state of S can be determined by k kind variable, be respectively x ₁... x _kIn real time figure is by behind the correlator, and the reference point that obtains after relevant with the storehouse image is got g reference point on the brightest reference point next door or times got g reference point, and its brightness is M ₁M _g, the storehouse image of establishing g reference point correspondence is Fig. 1 ... figure g, " N obtains relative reference point brightness m to general g after the normalization ₁M _g, necessarily satisfy equation

$F(x_1...x_k)=\left\{\begin{array}{c}{\mathrm{<figure-callout\; id="1"\; label="m"\; filenames="A200810227572E00152.png,A200810227572E00171.png"\; state="\{\{state\}\}">m</figure-callout>}}_1=f_1(x_1...x_k)\\ m_2=f_2(x_1...x_k)\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ m_g=f_g(x_1...x_k)\end{array}\right.$

Function f wherein ₁(x ₁... x _k) can by image 1 respectively with Fig. 1 ... the result who obtains after figure g (or with N width of cloth storehouse image) is relevant carries out curve fitting and obtains, same method, promptly use respectively Fig. 2, Fig. 3 ..., figure g and Fig. 1 ... the result who obtains after figure g (or with N width of cloth storehouse image) is relevant carries out curve fitting and can obtain f ₂(x ₁... x _k) ... f _g(x ₁... x _k).

Estimate rule two:

Function f ₁(x ₁... x _k), f ₂(x ₁... x _k) ... f _g(x ₁... x _k) between relation can be by Fig. 1 ... relation (variations such as translation, rotation, the convergent-divergent) decision of figure g supposes that promptly with m width of cloth figure be benchmark, f ₁(x ₁... x _k), f ₂(x ₁... x _k) ... f _g(x ₁... x _k) can use f _m(x ₁... x _k) represent h wherein ₁, h ₂... h _gBy Fig. 1 ... relation (variations such as translation, rotation, the convergent-divergent) decision of figure g then has

$\left\{\begin{array}{c}{f}_1(x_1...x_k)=h_1\left(f_m(x_1...x_k)\right)\\ f_2(x_1...x_k)=h_2\left(f_m(x_1...x_k)\right)\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ f_g(x_1...x_k)=h_g\left(f_m(x_1...x_k)\right)\end{array}\right.$

Estimate rule three:

According to estimating rule one and estimating rule two, provide complete estimate equation

$F(x_1...x_k)=\left\{\begin{array}{c}{m}_1{=h}_1\left(f_m(x_1...x_k)\right)\\ m_2=h_2\left(f_m(x_1...x_k)\right)\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ m_g=h_g\left(f_m(x_1...x_k)\right)\end{array}\right.$

K unknown number, g equation separated above-mentioned equation, obtains x ₁... x _kValue, like this, the state of S can be determined by the value of k kind variable.General g〉k, equation number can improve the precision of separating of equation after disturbance to a great extent more than the unknown number number.

In sum, volume holographic image identification multi-sample parallel estimation technique workflow is done following processing, multi-sample parallel estimation technique flow process is consulted accompanying drawing 2.

(6) with storehouse image 1 ... figure N deposits in the body holography correlator.

(7) will be used for the image 1/ of normalized ... figure N/ is input to respectively in the correlator, obtains image 1/ respectively ... figure N/ is with respect to storehouse image 1 ... the brightness of figure N reference point can be used as follow-up normalization benchmark.

(8) will scheme in real time in the S input body holography correlator, obtain itself and image 1 respectively ... normalized is carried out in the reference point brightness of figure N, obtains relative reference point brightness.

(9), carry out various parallel estimation and handle for the storehouse image.

4.1 determine predictor x ₁... x _kAnd predictor number k, determine sampling reference point number g;

4.2 according to estimating rule one, the estimate equation that provides estimation function and get in touch with reference point;

If be spaced apart L between the sampling reference point, the persistence length of correlation curve is Δ L (Δ L obtains according to estimation function),

Can carry out following estimation under g*L＜2 Δ L situations;

4.3 according to estimating rule two, get and decide the base estimation image, provide the relation between each estimation function;

4.4, provide complete estimate equation, and obtain the state difference of real-time figure with respect to the storehouse image by estimate equation according to estimating rule three.

(10) the above analysis can obtain the physical state that k ties up the real-time figure S of state.

According to above-mentioned 1-5 steps, on body holography correlator, adopt parallel estimation technique, its result's precision is compared with the common aspect holography correlator that does not use parallel estimation technique and is significantly increased.

Estimate rule by choosing suitable multisample, after estimating through multi-sample parallel, the state of figure can be determined by the reference point brightness value that obtains after relevant with the different sink image in real time.The present invention will change the drawback of classic method, change recognition result between the image of storehouse certain image from a certain storehouse image, play jamproof effect to a great extent simultaneously, thereby can significantly improve the holographic relevant matches precision of body.

Compared by experiment and do not used the multisample results estimated and use the multisample results estimated, can find that the output result after multi-sample parallel is estimated is more approaching with the virtual condition of real-time figure, so, can improve the matching precision of body holography correlator based on the holographic pretreated multi-sample parallel estimation technique of body.

Description of drawings

Fig. 1 is the principle schematic of body holography correlator.Wherein, 100 is input picture; 101 is Fourier transform lens; 102 is the body holographic storage material; 103 is the inversefouriertransform lens; 104 is the output face of body holography correlator; 105 for being stored in the storehouse image in 102.

Fig. 2 is a schematic flow sheet of the present invention.

Fig. 3 looks like to cut apart storehouse image synoptic diagram for embodiment 1 supergraph to be identified.

Fig. 4 is embodiment 1f ₁(x), f ₂(x) ... f ₄(x) matched curve.

Fig. 5 is embodiment 2 a supergraph picture to be identified.

Fig. 6 looks like to cut apart storehouse image synoptic diagram for embodiment 2 supergraphs to be identified.

Fig. 7 is embodiment 2f ₁(x, y), f ₂(x, y) ... f ₁₆(x, y) matched curve.

Embodiment

Be described in further detail below in conjunction with two specific embodiments and accompanying drawing specific implementation process the multi-sample parallel estimation technique that is applied to the body holography.

Embodiment one

Present embodiment is one and carries out the example that the multi-sample parallel technology improves precision to being applied in storehouse image on the body holography correlator and input picture.In the process of the holographic identification of body, the problem that one class translation identification is arranged, accompanying drawing 3 is a supergraph picture to be identified, wherein the image that yellow frame enclosed is designated as figure A, to scheme A is benchmark 10,20,30 pixels that move right respectively, form the image that red block, blue frame, purple frame are surrounded, be designated as B, C, D respectively; The image that black box is enclosed is the real-time figure of input, and the resolution of wherein every width of cloth image is 640 * 480.If do not adopt multi-sample parallel estimate to handle like this, Shi Bie result can only be an original image or to the image of 10 units of right translation, 20 units or 30 units, accuracy of identification is lower so.Adopt the flow process of multi-sample parallel estimation technique as follows: note original image is A, is respectively B, C, D to the image of 10 units of right translation, 20 units or 30 units, and real-time figure is S.

1, deposits image A, B, C, D in body holography correlator.

2, will scheme A ... figure D input body holography correlator obtains the brightness value of auto-correlation point.

3, will scheme in real time in the S input body holography correlator, obtain itself and storehouse image graph A respectively ... the reference point brightness of figure D is more respectively divided by storehouse image graph A ... the brightness of figure D auto-correlation point obtains relative reference point brightness and is respectively 0.801,0.939,0.735,0.514.

4, for the storehouse image A ... figure D carries out multi-sample parallel and estimates to handle

4.1 determine that sampling reference point number is 4;

4.2 according to estimating rule one, the estimate equation that provides estimation function and get in touch with reference point;

In the present embodiment, for 4 width of cloth storehouse image graph A that are stored in the body holography correlator ... figure D has arbitrarily figure S in real time, and the state of S is definite by a kind of variable x, i.e. transverse translation.Respectively near the brightest reference point (this for be the center) with bright spot 4 storehouse images of sampling are as sample point, after normalization, this 4 storehouse image pattern point brightness is 0.801,0.939,0.735,0.514

Necessarily satisfy equation

$\left\{\begin{array}{c}{f}_1\left(x\right)=0.801\\ f_2\left(x\right)=0.939\\ f_3\left(x\right)=0.735\\ f_4\left(x\right)=0.514\end{array}\right.$

Function f wherein ₁(x) can by image A respectively with image A ... the result who obtains after figure D is relevant carries out curve fitting and obtains, and as shown in Figure 4, establishing reference point brightness is n, then n=f (x, y)=[exp (α | x|)] ²Same method can obtain f ₂(x) ... f ₄(x), form and f ₁(x) identical, through over-fitting α ₁=0.01415, α ₂=0.01417, α ₃=0.01429, α ₄=0.01420 because the value difference of each function parameter α is little, calculate for convenience, get a mean value do approximate, α=0.01420.

Sampling reference point pixel at interval is 10, and persistence length is 1/ α=1/0.01420=70, and 4*10＜2*70 is then arranged, and can use multisample to estimate.

4.3 according to estimating rule two, get and decide the base estimation image, provide the relation between each estimation function;

Function f ₁(x), f ₂(x) ... f ₄(x) relation between can be by image A ... relation (transverse translation variation) decision of figure D,

To scheme A (leftmost diagram) is benchmark, f ₁(x), f ₂(x) ... f ₄(x) can use f ₁(x) represent, the picture of the leftmost side since 4 width of cloth figure, laterally the value of each storehouse image relative datum figure translational movement is respectively 0,10,20,30 unit, h so ₁, h ₂... h ₄By image A ... relation (transverse translation variation) decision of figure D, promptly

$\left\{\begin{array}{c}{f}_1\left(x\right)=h_1\left(f_1\left(x\right)\right)=f_1(x,)\\ f_2\left(x\right)=h_2\left(f_1\left(x\right)\right)=f_1(x-10)\\ f_3\left(x\right)=h_3\left(f_1\left(x\right)\right)=f_1(x-20)\\ f_4\left(x\right)=h_4\left(f_1\left(x\right)\right)=f_1(x-30)\end{array}\right.$

4.4, provide complete estimate equation, and obtain the state difference of real-time figure with respect to the storehouse image by estimate equation according to rule three.

Like this, we just can obtain complete estimate equation

$\left\{\begin{array}{c}{f}_1\left(x\right)=0.801\\ f_1(x-10)=0.939\\ f_1(x-20)=0.735\\ f_1(x-30)=0.514\end{array}\right.$

α=0.01420 wherein.

Separate above-mentioned equation like this, obtain x=8.15.

5. the above analysis according to the result of the variable x that obtains, is determined the state of figure S in real time.

According to The above results, the state that is to say real-time figure S is to be benchmark 8.15 pixels that move to right to scheme A.

And this test real-time figure virtual condition should be for being benchmark 8 pixels that move to right to scheme A, if do not use multi-sample parallel to estimate, and recognition result is for to scheme 10 pixels that move to right that A is a benchmark.

Utilize above-mentioned multi-sample parallel to estimate that ratio of precision had improved 93% originally like this.

Embodiment two

Present embodiment is the embodiment of a planar, and the state of figure will be determined by two physical quantitys in real time.In the process of image translation identification, accompanying drawing 5 is supergraph pictures to be identified, is designated as figure A; As shown in Figure 6, begin to be designated as first width of cloth figure from yellow frame, then since first width of cloth figure, every 5 pixels piece image that moves right, move 39 times altogether, form 40 width of cloth figure altogether, moving direction is an x direction among the figure, shown in blue frame, from above-mentioned 40 width of cloth figure, move piece image every 5 pixels downwards respectively, move 29 times altogether, moving direction is a y direction among the figure, shown in red block, so just formed one 40 * 30 two dimensional image storehouse, the resolution of wherein every width of cloth image is 640 * 480.At this moment import a width of cloth and scheme S in real time,, just can within 5 pixels, further discern, improved accuracy of identification to a great extent, but also can realize search in larger scope real-time figure if use the multi-sample parallel estimation technique.

Adopt the flow process of multi-sample parallel estimation technique as follows:

1. 1200 storehouse images are deposited in the body holography correlator.

2. 1200 inputs of storehouse image body holography correlator is obtained the brightness value of auto-correlation point, as follow-up normalization benchmark.

3. will scheme in real time in the S input body holography correlator, obtain the reference point brightness of itself and 1200 storehouse images respectively.

4,, carry out multi-sample parallel and estimate to handle for 1200 storehouse images

4.1 determine that this sub-sampling reference point number is 16, i.e. 4 * 4 on the two dimension.

4.2 according to estimating rule one, the estimate equation that provides estimation function and get in touch with reference point;

In the present embodiment, for 16 width of cloth storehouse images 1 that are stored in the body holography correlator ... Figure 16 has arbitrarily figure S in real time, and the state of S can be determined by 2 kinds of variablees, be respectively x, y, i.e. two-dimension translational.Respectively near the brightest reference point (this for be the center) with bright spot 4 * 4 storehouse images of sampling are as sample point, after normalization, this 4 * 4 storehouse image pattern point brightness is in order to descend 4 * 4 matrix representations

$\left[\begin{array}{cccc}0.392& 0.506& 0.581& 0.452\\ 0.551& 0.769& 0.833& 0.639\\ 0.517& 0.664& 0.769& 0.601\\ 0.364& 0.461& 0.544& 0.423\end{array}\right]$

Necessarily satisfy equation

$\left[\begin{array}{cccc}{f}_1(x,y)=0.392& f_2(x,y)=0.506& f_3(x,y)=0.581& f_4(x,y)=0.452\\ f_5(x,y)=0.551& f_6(x,y)=0.769& f_7(x,y)=0.833& f_8(x,y)=0.639\\ f_9(x,y)=0.517& f_{10}(x,y)=0.664& f_{11}(x,y)=0.769& f_{12}(x,y)=0.601\\ f_{13}(x,y)=0.364& f_{14}(x,y)=0.461& f_{15}(x,y)=0.544& f_{16}(x,y)=0.423\end{array}\right]$

Wherein function f 1 (x, y) can by image 1 respectively with image 1 ... the result who obtains after Figure 16 is relevant carries out curve fitting and obtains, and as shown in Figure 7, establishing reference point brightness is n, then n=f (x, y)=[exp ((α | x|+ β | y|)] ²Same method can obtain f ₂(x, y) ... f ₁₆(x, y), form and f1 (x, y) identical, on average obtain α=0.07 through over-fitting, β=0.05.

Sampling reference point pixel at interval is 5, and persistence length is 1/ α=1/0.07=14, and 1/ β=1/0.05=20 then has 4*5＜2*14,

4*5＜2*20 can use multisample to estimate.

4.3 according to estimating rule two, get and decide the base estimation image, provide the relation between each estimation function;

Function f ₁(x, y), f ₂(x, y) ... f ₁₆(x, y) relation between can be by image 1 ... the relation of image 16 (two-dimension translational variation) decision,

With the 1st width of cloth figure (upper left corner) is benchmark, f ₁(x, y), f ₂(x, y) ... f ₁₆(x y) can use f ₁(x y) represents, the picture in the upper left corner, the zone since 4 * 4, and laterally the value of each storehouse image relative datum figure is respectively 0,5,10,15 unit, and vertically the value of each storehouse image relative datum figure is respectively 0,5,10,15 unit, h ₁, h ₂... h ₁₆By image 1 ... the relation of image 16 (two-dimension translational variation) decision, promptly

$\left\{\begin{array}{c}{f}_1(x,y)=h_1\left(f_1(x,y)\right)=f_1(x,y)\\ f_2(x,y)=h_2\left(f_1(x,y)\right)=f_1(x-5,y)\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ f_{16}(x,y)=h_{16}\left(f_1(x,y)\right)=f_1(x-15,y-15)\end{array}\right.$

4.4, provide complete estimate equation, and obtain the state difference of real-time figure with respect to the storehouse image by estimate equation according to rule three.

Like this, we just can obtain complete estimate equation

$\left[\begin{array}{cccc}{f}_1(x,y)=0.392& f_1(x-5,y)=0.506& f_1(x-10,y)=0.581& f_1(x-15,y)=0.452\\ f_1(x,y-5)=0.551& f_1(x-5,y-5)=0.769& f_1(x-5,y-5)=0.833& f_1(x-15,y-5)=0.639\\ f_1(x,y-10)=0.517& f_1(x-5,y-10)=0.664& f_1(x-10,y-10)=0.769& f_1(x-15,y-10)=0.601\\ f_1(x,y-15)=0.364& f_1(x-5,y-10)=0.461& f_1(x-10,y-10)=0.544& f_1(x-15,y-15)=0.423\end{array}\right]$

α=0.07 wherein, β=0.05.

Separate above-mentioned equation like this, obtain x=7.1, y=8.9

5. the above analysis according to the result of the two-dimentional variable that obtains, is determined the state of figure S in real time.

According to The above results, the state that is to say real-time figure S is to be benchmark 7.1 pixels that move to right with upper left corner picture, moves down 8.9 pixels.

And this is tested real-time figure virtual condition and should move down 9 pixels for upper left corner picture being benchmark 7 pixels that move to right, if do not use multi-sample parallel to estimate, and recognition result moves down 10 pixels for upper left corner picture being benchmark 5 pixels that move to right.

Utilize above-mentioned multi-sample parallel to estimate that ratio of precision had improved 95% originally like this.

Claims (1)

1, improve the multi-sample parallel estimation method of matching precision of holographic correlator, it is characterized in that,

If the storehouse image of having stored is the N width of cloth, definition N width of cloth storehouse image is respectively Fig. 1～figure N, schemes S in real time, and the breadth of every width of cloth figure is P * Q, and wherein, P, Q are pixel count, and the state of S is determined by k physical quantity, is respectively x ₁～x _kFigure carries out related calculation with the storehouse image and obtains reference point in real time, and number is designated as N;

This method may further comprise the steps:

(1) storehouse image graph 1～figure N is deposited in the body holography correlator;

(2) will be used for the image 1 of normalized ^/～figure N ^/Be input to respectively in the correlator, obtain image 1 respectively ^/～figure N ^/With respect to storehouse image 1～figure N reference point brightness, can be used as follow-up normalization benchmark;

(3) will scheme in real time in the S input body holography correlator, obtain the reference point brightness of itself and image 1～figure N respectively, carry out normalized, obtain relative reference point brightness;

(4), carry out various parallel estimation and handle for the storehouse image;

(4.1) determine predictor x ₁～x _kAnd predictor number k, determine sampling reference point number g;

(4.2) according to estimating rule one, the estimate equation that provides estimation function and get in touch with reference point;

If be spaced apart L between the sampling reference point, the persistence length of correlation curve is Δ L, and Δ L obtains according to estimation function, carries out following estimation under g*L＜2 Δ L situations;

Described estimation rule one is:

In real time figure is by behind the correlator, and the reference point that obtains after relevant with the storehouse image is got g reference point on the brightest reference point next door or times got g reference point, and its brightness is M ₁～M _g, the storehouse image of establishing g reference point correspondence is Fig. 1～figure g, " N obtains relative reference point brightness m to g after the normalization ₁～m _g, necessarily satisfy equation

$F(x_1...x_k)=\left\{\begin{array}{c}{m}_1=f_1(x_1...x_k)\\ m_2=f_2(x_1...x_k)\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ m_g=f_g(x_1...x_k)\end{array}\right.$

Function f wherein ₁(x ₁... x _k) can by image 1 respectively with Fig. 1～figure g or relevant with N width of cloth storehouse image after the result that obtains carry out curve fitting and obtain, same method, promptly use respectively Fig. 2, Fig. 3 ..., figure g and Fig. 1 ... figure g or relevant with N width of cloth storehouse image after the result that obtains carry out curve fitting and can obtain f ₂(x ₁... x _k) ... f _g(x ₁... x _k);

(4.3) according to estimating rule two, get and decide the base estimation image, provide the relation between each estimation function;

Described estimation rule two is:

Function f ₁(x ₁... x _k), f ₂(x ₁... x _k) ... f _g(x ₁... x _k) between relation by Fig. 1 ... the relation decision of figure g supposes that with m width of cloth figure be benchmark, f ₁(x ₁... x _k), f ₂(x ₁... x _k) ... f _g(x ₁... x _k) all use f _m(x ₁... x _k) represent h wherein ₁, h ₂... h _gBy Fig. 1 ... the relation decision of figure g then has

$\left\{\begin{array}{c}{f}_1(x_1...x_k)=h_1\left(f_m(x_1...x_k)\right)\\ f_2(x_1...x_k)=h_2\left(f_m(x_1...x_k)\right)\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ f_g(x_1...x_k)=h_g\left(f_m(x_1...x_k)\right)\end{array}\right.$

(4.4) according to estimating rule three, provide complete estimate equation, and obtain the state difference of real-time figure with respect to the storehouse image by estimate equation;

Described estimation rule three is:

According to estimating rule one and estimating rule two, obtain complete estimate equation

$F(x_1...x_k)=\left\{\begin{array}{c}{m}_1=h_1\left(f_m(x_1...x_k)\right)\\ m_2=h_2\left(f_m(x_1...x_k)\right)\\ \&CenterDot;\&CenterDot;\&CenterDot;\\ m_g=h_g\left(f_m(x_1...x_k)\right)\end{array}\right.$

K unknown number, g equation separated above-mentioned equation, obtains x ₁～x _kValue, like this, the state of S is determined by the value of k kind variable; Work as g〉k, equation number can improve the precision of separating of equation after disturbance more than the unknown number number;

(5) obtain the physical state that k ties up the real-time figure S of state.

Images