CN101666682B - Neural network nonuniformity correction method based on scene statistics - Google Patents

Abstract

The invention requests to protect a neural network nonuniformity correction method based on scene statistics, belonging to the infrared focal plane detection field. The invention provides the method aiming at the shortcoming that the traditional neural network correction method is difficult to eliminate low-frequency spatial noises. The method comprises the following steps: initializing image related matrices and parameters; detecting and compensating blind pixels according to the image pixel gray value; carrying out nonuniformity correction on image offset by a scene statistics method; judging the regional attributes of the pixels by a neural network correction method according to the standard deviation thresholds of correction errors, and carrying out nonlinear gain correction on the images which are corrected based on scene statistics and contain no low-frequency spatial noises. By the method, the correction effects of the desired image signals are good, target fade-out and ghosting are inhibited, and the changes of the background images hardly affect the correction effects. The method can be widely applied to image detection and processing.

Description

Neural network nonuniformity correction method based on scene statistics

Technical field

The present invention relates to the image detection and processing technical field, specifically belong to method for correcting image in the infrared focus plane Detection Techniques.

Background technology

The infrared focal plane array image-forming system is owing to have highly sensitive; Volume is little; Compact conformation; Operating distance is far away, anti-interference good, penetrate the smog ability strong, can be round-the-clock, advantage such as round-the-clock work, become the infrared imagery technique Development Trend, and the gazing type infrared focal plane array has become the main flow sensitive detection parts of following infra-red thermal imaging system development.But owing to receive the restriction of material and technological level; Ubiquity heterogeneity between each probe unit response characteristic of infrared focal plane array (IRFPA); It will cause the performances such as temperature resolution of infrared imaging system significantly to descend; So that make it be difficult to satisfy the practical applications requirement, thereby the infrared focal plane array that uses in the engineering almost all adopts the Nonuniformity Correction technology without any exception.

Diversified infrared focal plane array heterogeneity bearing calibration has appearred at present both at home and abroad; But conclude to get up roughly can be divided into two types: one type of bearing calibration that is based on demarcation mainly comprises 2 temperature calibration methods (TPC) and multi-point temp standardization (ETPC).Such bearing calibration has the algorithm simple and flexible, and fast operation is easy to advantages such as hardware realization, is the main method of present practical applications.But owing to receive the influence of infrared focal plane array working time and environmental change, slow drift can take place in its response parameter, and then the effect correction precision, therefore, demarcates a type bearing calibration and need carry out periodicity demarcation correction usually.Like this, not only need interrupt the real time imagery process, and complicated operation.The another kind of bearing calibration that is based on scene mainly comprises constant statistical average method (CSC), time domain high-pass filtering method (THPFC) and artificial neural network method (ANNC).These class methods can be eliminated infrared focal plane array effectively with the response parameter drift that working time and environmental change take place, and do not need calibration, only need to proofread and correct according to scene information realization IRFPA heterogeneity self-adaptive.In bearing calibration, the most representative with the neural network bearing calibration based on scene.

The tradition neural network correction method (Scribner) though need not calibrate in theory fully infrared focal plane array, not high to the linearity and the stability requirement of parameter detector yet.But also there is tangible deficiency in traditional neural network correction method, and is particularly powerless to the low frequency space noise.Trace it to its cause be simply with neighbours territory average as desired output, the pixel of participate in calculating is not particularly considered current pixel value very little, causes expectation value and actual value possibly have bigger deviation.Signal changes slowly in the same area in the scene, and neighbours territory mean value computation expectation value is rational; When spatially there was acute variation in signal, at this moment window just existed bigger error with neighbours territory average as expectation value across a plurality of different zones.When target travel, this error can not accumulated, and is not obvious to the calibration result influence; When target was tending towards static for a long time, this error is accumulation rapidly, and iteration step length is big more, and error accumulation is fast more, then tangible target fade-out can occur.And become the correction diplopia that next reverted image will be in situ stayed the time in unexpected motion by static when target.

People such as Zhang Tianxu, Shi Yan is in " spatial frequency characteristic of infrared focal plane asymmetric noise and spatially adaptive asymmetric correction method improve " literary composition; Analyzed the spatial frequency characteristic of infrared focal plane array heterogeneity noise, pointed out that the space low-frequency noise is its principal ingredient.The deficiency that the low frequency space noise exists is removed in space domain self-adapted bearing calibration to tradition, proposes the method that adopts a point calibration and space domain self-adapted correction to combine.Point calibration in this method is to obtain through continuous limited frame and the background that do not contain target are carried out time average.If when camera duration of work image background remains unchanged, time average obtains background image and real background is approaching through the limited frame background that does not contain target is continuously carried out, and at this moment can obtain calibration result preferably.But under the situation of background image real-time change, there are bigger error in background image that obtains according to the method and real background, and the result of pre-service correction will inevitably influence final correcting result like this.

Existence owing to the infrared focal plane array blind element causes infrared image smudgy, and mainly is to detect blind element through signal processing technology at present, and effectively compensates to improve the infrared focal plane array image-forming quality through neighborhood territory pixel.The processing of blind element comprises that blind element detects and compensate two aspects.Traditional blind element disposal route is difficult to realize the online detection and the compensation of blind element.In Dai Shaosheng, Zhang Tianqi " a kind of new infrared focal plane array blind element Processing Algorithm " literary composition, the instant new algorithm that detects and compensate of a kind of IRFPA blind element has been proposed.This algorithm is realized simple, and highly versatile can detect and compensate the blind element that produces at random immediately, but only relates to the detection and the compensation of blind element, does not relate to picture signal that Nonuniformity Correction is carried out in biasing and nonlinear gain is proofreaied and correct.

The suffered spatial noise of infrared focal plane array has following two attributes:

Character one: it mainly is rendered as low-frequency component the spatial noise that is caused by heterogeneity.

Character two: the spatial noise by the gain heterogeneity causes separately mainly is rendered as radio-frequency component.Yet the prerequisite of traditional neural network bearing calibration hypothesis is: the spatial noise that heterogeneity caused, its spatial frequency mainly are high frequency or white noise.Just be based on this prerequisite, the traditional neural networks bearing calibration adopts 4 neighborhood averages of pixel to upgrade correction coefficient as the correction expectation value of this pixel output, makes the frequency characteristic of its presentation space low pass.Therefore, when actual infrared focal plane array spatial noise when being main with low frequency, the traditional neural networks bearing calibration seems powerless.

The algorithm (onepoint_nn_nuc) that combines like a point calibration and space domain self-adapted correction; The variation meeting of its background image exerts an influence to the effect of proofreading and correct; It is better that background changes less part calibration result relatively, and the part calibration result that change of background is bigger is relatively poor.

If we adopt certain pre-service to proofread and correct in advance, eliminate space low frequency part noise, only remaining by the gain spatial high-frequency noise that heterogeneity caused, and then adopt neural net method to carry out follow-up correction, just can obtain calibration result preferably.Just for these reasons; The present invention proposes neural network nonuniformity correction method based on scene statistics; Promptly at first eliminate the space low-frequency noise that causes by IRFPA biasing heterogeneity, and then adopt neural net method to carry out follow-up correction through scene statistics.

Summary of the invention

The present invention is directed to and cause infrared focal plane array response characteristic and change of stability under the environmental baseline complicated and changeable; And traditional neural network bearing calibration is difficult to eliminate the deficiency of low frequency space noise, proposes a kind of based on scene statistics and neural network nonuniformity correction method.This method comprises, carries out that blind element detects and compensation, adopts the scene statistics method to image biasing carrying out Nonuniformity Correction, adopts the neural network correction method that scene statistics is proofreaied and correct the back and the image that do not contain the low frequency space noise carries out nonlinear gain and proofreaies and correct.

Blind element detects and compensation.Before infrared focal plane array heterogeneity is proofreaied and correct, at first need carry out blind element and detect and compensation.With the time domain average of continuous k frame image sequence responsiveness B as each pixel of present frame _Ij, promptly $B_{\mathrm{Ij}}=\frac{\underset{t=n-(k-1)}{\overset{n}{\mathrm{\Σ}}}{X}_{i,j}\left(t\right)}{k};$ Judge pixel B _IjWhether be effective pixel, at first with pixel B _IjFor the center size is the minimum and maximum grey scale pixel value B of the interior inquiry of window of (2h+1) * (2h+1) _Max, B _MinRemove B then _Max, B _Min, residual pixel average gray B in the calculation window, promptly $\stackrel{\‾}{B}=\frac{\underset{i=h}{\overset{i+h}{\mathrm{\Σ}}}\underset{j=h}{\overset{j+h}{\mathrm{\Σ}}}{B}_{\mathrm{Ij}}-B_{\mathrm{Max}}-B_{\mathrm{Min}}}{(2h+1)\×(2h+1)-2};$ Each pixel is carried out blind element judge, when satisfying condition $\mathrm{\Δ}=\frac{B_{\mathrm{Max}}-\stackrel{\‾}{B}}{\stackrel{\‾}{B}}\≥9$ Perhaps $\mathrm{\Δ}=\frac{\stackrel{\‾}{B}-B_{\mathrm{Min}}}{B_{\mathrm{Min}}}\≥9$ The time, decidable is a blind element, otherwise is valid pixel.If pixel B _IjBe blind element, then relevant position in the blind element matrix (i.e. the capable j row of i) put 1; Detected blind element position in the current frame original image is replaced compensation with blind element pixel 4 neighborhood averages, obtain removing the image of blind element.

Adopt the scene statistics method to image biasing carrying out Nonuniformity Correction.Time domain mean value computation original image input average E [X according to original image _{I, j}], and proofread and correct output average E [Y _{I, j}], do not comprised the gray-scale value Z of each pixel of low frequency space noise _{I, j}(n).

Utilize the 1st frame to of the input of n frame original image time domain average, can adopt recursive mode to calculate E [X as infrared focal plane array _{I, j}]: promptly $E\left[X_{i,j}\right]=\stackrel{\‾}{X_{i,j}}=\frac{X_{i,j,n}+(n-1)\·X_{i,j,n-1}}{n}.$ Can adopt X _{I, j}Space average as the average output E [Y of infrared focal plane array n frame pixel _{I, j}]: $E\left[Y_{i,j}\right]=\stackrel{\‾}{X}=\frac{1}{M\×N}\×\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}\stackrel{\‾}{X_{i,j}}.$

The original image X of present frame _{I, j}(n) deduct infrared focal plane array (i, j) pixel average input X in time _{I, j}, just can not contained each grey scale pixel value Z of low frequency space noise _{I, j}(n).

Z _i，j(n)＝X _i，j(n)-X _i，j

Adopt the neural network correction method that scene statistics is proofreaied and correct, the image that does not contain the low frequency space noise is carried out nonlinear gain proofread and correct.As adopt the self-adaptive weighted average wave filter, confirm the wanted signal of output.Judge the area attribute of pixel according to the standard deviation threshold method of correction error, the pixel of the same area is distributed bigger weights, the pixel of zones of different is distributed less weights, by the weights W of weighted mean wave filter _{P, q}(n) the expectation value F of definite output _{I, j}(n).

Wherein, weights are confirmed by following formula:

$W_{\mathrm{pq}}\left(n\right)=\frac{1}{|Z_{p,q}\left(n\right)-Z_{i,j}\left(n\right)|\&CenterDot;\mathrm{\&eta;}+1}\&CenterDot;\left(\right|Z_{p,q}\left(n\right)-Z_{i,j}\left(n\right)|<\mathrm{th})$

Wherein, η is the weights coefficient.

The output of weighting filter is expectation value F _{I, j}(n), according to computes expectation value F _{I, j}(n):

$F_{i,j}\left(n\right)=\frac{\underset{p=i-h}{\overset{i+h}{\mathrm{\&Sigma;}}}\underset{q=j-h}{\overset{j+h}{\mathrm{\&Sigma;}}}{W}_{p,q}\left(n\right)Z_{p,q}\left(n\right)}{\underset{p=i-h}{\overset{i+h}{\mathrm{\&Sigma;}}}\underset{q=j-h}{\overset{j+h}{\mathrm{\&Sigma;}}}{W}_{p,q}\left(n\right)}$

The present invention compares with the method that a point calibration and neural network combine with traditional neural net method, has stronger Nonuniformity Correction ability, thereby can obtain desirable image rectification effect.In the scene continually varying is proofreaied and correct in real time, can eliminate the low frequency space noise, can eliminate target fade-out and ghost again, improved the arithmetic speed of algorithm.

Description of drawings

Fig. 1 is a bearing calibration schematic flow sheet of the present invention.

Fig. 2 detects synoptic diagram for blind element.Wherein, Fig. 2 (a) is a k frame moving scene synoptic diagram; Fig. 2 (b) is 3 * 3 blind element detection windows.

Embodiment

The implementation procedure of analytical algorithm on the principle at first below:

Adopt linear model to proofread and correct to infrared focal plane array, calibration model is:

Y _i，j(n)＝G _i，j(n)·X _i，j(n)+O _i，j(n) (1)

Y wherein _{I, j}Output, X are proofreaied and correct in expression _{I, j}The input of expression original image, G _{I, j}The expression correcting gain, O _{I, j}The biasing of expression correcting image.

The expression formula that expectation can obtain following form is got on the following formula both sides:

E[Y _i，j]＝G _i，j(n)·E[X _i，j]+O _i，j(n) (2)

Wherein E [] is an expectation operator, E [Y _{I, j}] for proofreading and correct output average, E [X _{I, j}] be that original image is imported average,

Obtain following formula after (1) formula and (2) formula subtracted each other:

Y _i，j(n)-E[Y _i，j]＝G _i，j(n)·(X _i，j(n)-E[X _i，j]) (3)

Formula (3) is compared and can be found out with formula (1), and formula has increased E [X in (3) _{I, j}], lacked bias term O _{I, j}With in the former algorithm to the biasing O _{I, j}Calculating estimate to be converted into [X to input signal average E _{I, j}] and proofread and correct output average E [Y _{I, j}] calculating.If make z _{I, j}(n)=Y _{I, j}(n)-E [Y _{I, j}], Z _{I, j}(n)=X _{I, j}(n)-E [X _{I, j}] can get so:

z _i，j(n)＝G _i，j(n)·Z _i，j(n) (4)

(4) no longer contain the spatial noise of low frequency in the formula, can proofread and correct the high frequency spatial noise that is caused by the gain heterogeneity with neural net method this moment again.Correction is output as:

Y _i，j(n)＝G _i，j(n)·(X _i，j(n)-E[X _i，j])+E[Y _i，j] (5)

Below input signal average E [X further is discussed _{I, j}] and proofread and correct output average E [Y _{I, j}] calculating.

Because the temporal correlation of image sequence, the gray average of n frame original image is as infrared focal plane array time domain average input value before utilizing.Use X _{I, j, k}Represent the K two field picture (i, j) the pixel original value is used X _{I, j}The expression infrared focal plane array the (i, the j) time average of pixel input, then: $\stackrel{\&OverBar;}{X_{i,j}}=\frac{\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{X}_{i,j,k}}{n}.$ Use X _{I, j}Replace E [X _{I, j}], obtain following formula:

$E\left[X_{i,j}\right]=\stackrel{\&OverBar;}{X_{i,j}}=\frac{\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{X}_{i,j,k}}{n}---\left(6\right)$

Because the X of original image _{I, j}Correlativity spatially will constantly strengthen along with the increase of accumulation frame number, therefore can use X _{I, j}Spatial domain average X replace (i, j) E [Y on average exported in the correction of pixel _{I, j}]: promptly

$E\left[Y_{i,j}\right]=\stackrel{\&OverBar;}{X}=\frac{1}{M\&times;N}\&times;\underset{i=1}{\overset{R}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{L}{\mathrm{\&Sigma;}}}\stackrel{\&OverBar;}{X_{i,j}}---\left(7\right)$

M, N are focal plane arrays (FPA) line number and columns.

(6) formula and (7) formula substitution (5) formula are obtained final correction to be output as:

Y _i，j(n)＝G _i，j(n)·(X _i，j(n)-X _i，j)+X (8)

X wherein _{I, j}Be respectively (6) formula and (7) formula result calculated with X.

Below to accompanying drawing and instance enforcement of the present invention is specifically described, Fig. 1 is a bearing calibration schematic flow sheet of the present invention, specifically may further comprise the steps: initialization, blind element detection and compensation, scene statistics and neural network are proofreaied and correct.

(1) initialization procedure

At first carry out initialization, initialisation image two-dimensional matrix and parameter.The gain correction coefficient G of each pixel of initialization is complete 1 matrix; The blind element storage matrix is complete 0 matrix; It is complete 0 matrix that time domain is imported equal value matrix, and spatial domain output average is 0, original sequence frame=1 to be corrected; Error threshold th=0.2, setting the original image frame number k that is used for the blind element detection computations is natural number.

(2) blind element detects and compensation

According to the responsiveness of image pixel gray-scale value calculating pixel, be the center with the responsiveness, inquire about confirming the pixel grey scale average in the window, find out minimum and maximum grey scale pixel value B _Max, B _MinRemove B _Max, B _Min, residual pixel average gray B in the calculation window confirms the position of blind element thus, and the pixel of blind element matrix relevant position is put 1; Detected blind element position is replaced compensation with blind element pixel 4 neighborhood averages, obtain removing the image of blind element.

The present invention adopts the real-time blind element detection algorithm based on scene that the blind element at random that produces in the infrared focal plane array course of work is detected immediately.Adopt present frame and continuous k-1 two field picture before thereof, the gray-scale value of each pixel is carried out time domain average find the solution its responsiveness, and be designated as B _IjB then _IjCan be expressed as: $B_{\mathrm{Ij}}=\frac{\underset{t=n-k+1}{\overset{n}{\mathrm{\&Sigma;}}}{X}_{i,j}\left(t\right)}{k}$

Be illustrated in figure 2 as blind element and detect synoptic diagram.Carry out the responsiveness of time domain average with comprising present frame at interior continuous 10 frames (from the n-9 frame to the n frame) original image pixels gray-scale value among the figure as respective pixel.In order to improve operation efficiency, we adopt alternative manner to calculate B _Ij:

SumV _i，j(n)＝SumV _i，j(n-1)+X _i，j(n)-X _i，j(n-9)

$B_{\mathrm{ij}}=\frac{\mathrm{Sum}{V}_{i,j}\left(n\right)}{10}$

Utilize real-time blind element detection algorithm that the responsiveness that obtains is carried out the blind element detection again, and the element of blind element matrix relevant position is changed to 1 based on scene.

Its blind element testing process is following:

(a) with responsiveness B _IjBe the center, the pixel grey scale average in (2h+1) * (2h+1) window is inquired about, find out minimum and maximum grey scale pixel value B _Max, B _Min

(b) in window, remove B _Max, B _Min, and obtain the mean value B of residual pixel gray scale in the window, promptly $\stackrel{\&OverBar;}{B}=\frac{\underset{i=h}{\overset{i+h}{\mathrm{\&Sigma;}}}\underset{j=h}{\overset{j+h}{\mathrm{\&Sigma;}}}{B}_{\mathrm{Ij}}-B_{\mathrm{Max}}-B_{\mathrm{Min}}}{(2h+1)\&times;(2h+1)-2}$

(c) compare B _Max, B _MinWith the number percent of B difference, even $\mathrm{\&Delta;}=\frac{B_{\mathrm{Max}}-\stackrel{\&OverBar;}{B}}{\stackrel{\&OverBar;}{B}}$ Or $\mathrm{\&Delta;}=\frac{\stackrel{\&OverBar;}{B}-B_{\mathrm{Min}}}{B_{\mathrm{Min}}}.$ According to regulation among the GB GB/T1744421998 " infrared focus plane Acceptance Test technical standard ", when Δ>=9, think that then this pixel is a blind element, and note the position of blind element, the element of relevant position in the blind element matrix is changed to 1.

(d) blind element compensation.Detected blind element position in the current frame original image is replaced compensation with blind element pixel 4 neighborhood averages, obtain removing the image of blind element.

$X_{i,j}=\frac{X_{i-1,j}+X_{i+1,j}+X_{i,j-1}+X_{i,j+1}}{4}$

(2) scene statistics step

Adopt the scene statistics method that image is setovered and carry out Nonuniformity Correction, its objective is time domain mean value computation original image input average E [X according to original image _{I, j}], and proofread and correct output average E [Y _{I, j}], do not comprised the gray-scale value Z of each pixel of low frequency space noise _{I, j}(n).

Because image sequence life period correlativity, the time domain average of n frame original image gray-scale value is as the average input of infrared focal plane array (IRFPA) before utilizing.(i, j) pixel average input is in time provided by formula (6) infrared focal plane array.

In order to improve operation efficiency and to take less storage space, the present invention adopts a kind of recursive mode to calculate original image input average E [X _{I, j}], that is:

$E\left[X_{i,j}\right]=\stackrel{\&OverBar;}{X_{i,j}}=\frac{X_{i,j,n}+(n-1)\&CenterDot;X_{i,j,n-1}}{n}---\left(9\right)$

Output average E [Y is promptly proofreaied and correct in the average output that can be calculated the infrared focal plane array pixel by image correlativity spatially _{I, j}]:

$E\left[Y_{i,j}\right]=\stackrel{\&OverBar;}{X}=\frac{1}{M\&times;N}\&times;\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}\stackrel{\&OverBar;}{X_{i,j}}---\left(10\right)$

The original image X of present frame _{I, j}(n) deduct infrared focal plane array (i, j) pixel average input X in time _{I, j}(original image input average E [X _{I, j}]), just can not contained each grey scale pixel value Z of low frequency space noise _{I, j}(n).

Z _i，j(n)＝X _i，j(n)-X _i，j

(4) neural network is proofreaied and correct

The image that adopts the neural network correction method that scene statistics is proofreaied and correct the back and do not contain the low frequency space noise carries out nonlinear gain and proofreaies and correct.

Traditional neural network correcting algorithm that people such as D.A.Scribner propose just simply utilizes 4 neighborhood averages of current pixel to calculate its expectation value F _{I, j}(n).This method is inner to scene the same area, and the expectation value of calculating current pixel is reasonably, but at the edges of regions place, it is obviously not enough to utilize this method calculation expectation value to exist.Simultaneously, when using the expectation value of neighbours territory mean value computation current pixel,, can not obtain the most approaching real expectation value to the big IRFPA of space low-frequency noise because the pixel count that participation is calculated is few.For this reason; The present invention adopts the self-adaptive weighted average wave filter; Use more pixel to participate in average calculating operation; Judge the zone of pixel according to the standard deviation threshold method of correction error,, the pixel of zones of different is distributed less weights distributing bigger weights with the pixel of center pixel (the capable j row of i pixel) the same area.

For the interior pixel p q of (2h+1) filter window that with the ij pixel is the center, its weights are confirmed by following formula:

$W_{\mathrm{pq}}\left(n\right)=\frac{1}{|Z_{p,q}\left(n\right)-Z_{i,j}\left(n\right)|\&CenterDot;\mathrm{\&eta;}+1}\&CenterDot;({|Z}_{p,q}\left(n\right)-Z_{i,j}\left(n\right)|<\mathrm{th})$

η is the weights coefficient.

The output of weighting filter is expectation value F _{I, j}(n):

$F_{i,j}\left(n\right)=\frac{\underset{p=i-h}{\overset{i+h}{\mathrm{\&Sigma;}}}\underset{q=j-h}{\overset{j+h}{\mathrm{\&Sigma;}}}{W}_{p,q}\left(n\right)Z_{p,q}\left(n\right)}{\underset{p=i-h}{\overset{i+h}{\mathrm{\&Sigma;}}}\underset{q=j-h}{\overset{j+h}{\mathrm{\&Sigma;}}}{W}_{p,q}\left(n\right)}$

Try to achieve expectation value F _{I, j}(n) after, utilize following formula to pursue the frame iteration correction again:

Y _i，j(n)＝G _i，j(n)·Z _i，j(n)+X

G _i，j(n)＝G _i，j(n-1)-2λZ _i，j(n)·(Z _i，j(n)-F _i，j(n))

Y wherein _{I, j}(n) be that the n frame is proofreaied and correct output, G _{I, j}(n) be n frame gain correction coefficient, λ is an iteration step length.

For with pixel i, j is the pixel p q in 3 * 3 filter windows at center, and its 4 neighborhood weights (W1 represent pixel i-1, the weights of j deserve to be called weights; W2 represent pixel i+1, the weights of j are claimed weights down; W3 represent pixel i, the weights of j-1 are claimed Zuoquan's value; W4 represent pixel i, the weights of j+1 are claimed right weights) confirm by following formula:

$\mathrm{<figure-callout\; id="1"\; label="W"\; filenames="H2009101045511E00012.png"\; state="\{\{state\}\}">W</figure-callout>}1=\frac{1}{|Z_{i-1,j}\left(n\right)-Z_{i,j}\left(n\right)|\&CenterDot;\mathrm{\&eta;}+1}\&CenterDot;\left(\right|Z_{i-1,j}\left(n\right)-Z_{i,j}\left(n\right)|<\mathrm{th})$

$\mathrm{<figure-callout\; id="2"\; label="W"\; filenames="H2009101045511E00012.png"\; state="\{\{state\}\}">W</figure-callout>}2=\frac{1}{|Z_{i+1,j}\left(n\right)-Z_{i,j}\left(n\right)|\&CenterDot;\mathrm{\&eta;}+1}\&CenterDot;\left(\right|Z_{i+1,j}\left(n\right)-Z_{i,j}\left(n\right)|<\mathrm{th})$

$W3=\frac{1}{|Z_{i,j-1}\left(n\right)-Z_{i,j}\left(n\right)|\&CenterDot;\mathrm{\&eta;}+1}\&CenterDot;\left(\right|Z_{i,j-1}\left(n\right)-Z_{i,j}\left(n\right)|<\mathrm{th})$

$W4=\frac{1}{|Z_{i,j+1}\left(n\right)-Z_{i,j}\left(n\right)|\&CenterDot;\mathrm{\&eta;}+1}\&CenterDot;\left(\right|Z_{i,j+1}\left(n\right)-Z_{i,j}\left(n\right)|<\mathrm{th})$

η is the weights coefficient.

The output of weighting filter is expectation value F _{I, j}(n):

$F_{i,j}\left(n\right)=\frac{\mathrm{<figure-callout\; id="1"\; label="W"\; filenames="H2009101045511E00012.png"\; state="\{\{state\}\}">W</figure-callout>}1\&CenterDot;Z_{i-1,j}\left(n\right)+\mathrm{<figure-callout\; id="2"\; label="W"\; filenames="H2009101045511E00012.png"\; state="\{\{state\}\}">W</figure-callout>}2\&CenterDot;Z_{i+1,j}\left(n\right)+W3\&CenterDot;Z_{i,j-1}\left(n\right)+W4\&CenterDot;Z_{i,j+1}\left(n\right)+Z_{i,j}\left(n\right)}{\mathrm{<figure-callout\; id="1"\; label="W"\; filenames="H2009101045511E00012.png"\; state="\{\{state\}\}">W</figure-callout>}1+\mathrm{<figure-callout\; id="2"\; label="W"\; filenames="H2009101045511E00012.png"\; state="\{\{state\}\}">W</figure-callout>}2+W3+W4+1}$

Try to achieve expectation value F _{I, j}(n) after, utilize following formula to pursue the frame iteration correction again:

Y _i，j(n)＝G _i，j(n)·Z _i，j(n)+X

G _i，j(n)＝G _i，j(n-1)-2λZ _i，j(n)·(Z _i，j(n)-F _i，j(n))

Y wherein _{I, j}(n) be that the n frame is proofreaied and correct output, G _{I, j}(n) be n frame gain correction coefficient, λ is an iteration step length.

Original image calculation expectation signal is adopted in the traditional neural networks bearing calibration; Adopt during neural network of the present invention is proofreaied and correct and proofread and correct output image calculation expectation signal; Be used for the expectation estimation but also increased current pixel; And adopt the method calculation expectation value of adaptive weighted filter, thereby the present invention can overcome traditional neural network effectively and proofreaies and correct defectives such as the target fade-out that exists, ghost.The method that combines with a point calibration and neural network based on the scene statistics method that correction combines with neural network that the present invention proposes can both be eliminated the space low-frequency noise well.But when image background constantly changes; The method that one point calibration combines with neural network will there are differences because of initial sampled background image and movement background image and make the calibration result variation; When this error increased, correction error also increased thereupon, even target can be submerged in the background.The present invention is based on the bearing calibration that scene statistics and neural network combine, is to utilize the scene correlativity constantly to add up, in real time background image updating.In the scene continually varying was proofreaied and correct in real time, method calibration result of the present invention was good, has suppressed target fade-out and ghost, and the variation of background image is to almost not influence of calibration result.

Claims (4)

1. the neural network nonuniformity correction method based on scene statistics is characterized in that, specifically may further comprise the steps: initialisation image correlation matrix and parameter; Carrying out blind element according to the image pixel gray-scale value detects and compensates; Time domain mean value computation original image input average E [X according to original image _{I, j}], by the correction output average E [Y of image correlation calculations image pixel spatially _{I, j}], with the original image X of present frame _{I, j}(n) deduct original image input average E [X _{I, j}], do not comprised the image pixel gray-scale value Z of low frequency space noise _{I, j}(n), to image biasing carrying out Nonuniformity Correction; Adopt the neural network correction method; Judge the area attribute of image pixel according to the standard deviation threshold method of correction error, the image that scene statistics is proofreaied and correct the back and do not contained the low frequency space noise carries out nonlinear gain to be proofreaied and correct, and the area attribute of said judgement image pixel is specially; Adopt the self-adaptive weighted average wave filter; Pixel to the same area is distributed bigger weights, the pixel of zones of different is distributed less weights, by the weights W of weighted mean wave filter _{P, q}(n) the image expectation value F of definite output _{I, j}(n).

2. neural network nonuniformity correction method according to claim 1 is characterized in that, said blind element detects and compensation specifically comprises: according to image pixel gray-scale value calculating pixel responsiveness; With the responsiveness is the center, searches minimum and maximum image pixel gray average B _Max, B _MinRemove B _Max, B _Min, residual image pixel grey scale mean value in the calculation window again

Each pixel is carried out blind element judge, confirm the position of blind element thus, and the image pixel of relevant position in the blind element matrix is put 1; At last, detected blind element position is replaced compensation with blind element pixel 4 neighborhood averages, obtain removing the image of blind element.

3. neural network nonuniformity correction method according to claim 2; It is characterized in that; The concrete grammar of confirming the blind element position is: when satisfying condition

perhaps when

; This position is a blind element, otherwise is valid pixel.

4. neural network nonuniformity correction method according to claim 2 is characterized in that, the concrete steps of said calculating pixel responsiveness are: with the time domain average of continuous k frame image sequence grey scale pixel value as each pixel response rate B of present frame _Ij

Images