CN104599261A - Equalization histogram neural network heterogeneity correcting method for short wave infrared focal plane - Google Patents

Abstract

The invention relates to a neural network heterogeneity correcting method and a neural network heterogeneity correcting device for short wave infrared focal plane, wherein the method comprises the following steps: a time-domain statistic histogram counting step: performing time-domain statistic histogram counting on each pixel point in the short wave infrared focal plane; a neighborhood time-domain histogram mean value calculating step: calculating the equalization histogram of each pixel point by utilizing the time-domain histogram of the neighborhood pixel point of each pixel; a neutral network estimation value calculating step: calculating a grey estimation value of each pixel point for a neural network algorithm according to the mean-value histogram of each pixel point; a neutral network correcting step: correcting an original infrared image through ten neutral network algorithm to obtain heterogeneity corrected image output. According to the neural network heterogeneity correcting method and the neural network heterogeneity correcting device for the short wave infrared focal plane disclosed by the invention, the defect that a conventional neutral network correcting method needs a relatively movable image scene is improved, the algorithm is simple, the calculating complexity is low, and engineering practice is convenient.

Description

The equal value histogram neural network nonuniformity correcting algorithm of short-wave infrared focal plane

Technical field

The present invention relates to a kind of neural network nonuniformity correction method and device of short-wave infrared focal plane, belong to infrared image processing field.

Background technology

At present, have already been proposed many about nonuniformity correction (Nonuniformity Correction, NUC) algorithm. but generally speaking can be divided into two large classes: a class is the method (as peg method) based on demarcating, these class methods have higher calibration accuracy, but need many additional devices (black-body resource, optical device etc.), considerably increase volume and the cost of detector, and also must make imaging system break-off in calibration process, this seriously constrains the usable range of such algorithm.Another kind of is method based on scene, and it directly utilizes the scene information of every two field picture to carry out nonuniformity correction, overcomes the deficiency of first kind method, is therefore the focus of research at present.But most algorithm all needs to estimate real scene value in this kind of technology, typical as: Scribner etc. propose based on neural network (NeuralNetwork, NN) non-uniformity correction algorithm, this algorithm is using the estimated value of four neighborhood averagings of noise figure picture as true picture, and for neural network training, but prerequisite be Fixed-pattern noise can cut down by four neighborhood averagings, otherwise iteration will be dispersed; Hardie etc. propose based on average (the Motion-CompensatedTemporal Average of motion-compensated temporal, MCTA) correcting algorithm, image after this algorithm utilizes motion compensation average is to estimate real scene, but now must ensure that Same Scene will have fully many probe units to observe at different time, otherwise calibration result also can be undesirable.

Summary of the invention

The object of the invention is to, neural network nonuniformity correction method and the device of the short-wave infrared focal plane that a kind of algorithm is simple, computation complexity is low is provided.

In order to realize foregoing invention object, the invention provides following technical scheme:

A neural network nonuniformity correction method for short-wave infrared focal plane, is characterized in that, comprising:

Time domain statistics with histogram step: time domain statistics with histogram is carried out to each pixel in short-wave infrared focal plane;

Neighborhood time domain histogram mean value computation step: the equal value histogram of each pixel described in the time domain histogram calculation utilizing the neighborhood territory pixel of each pixel point;

The calculation procedure of neural network estimated value: the gray scale estimated value being used for each pixel of neural network algorithm according to the average histogram calculation of described each pixel;

Corrected neural network step: utilize neural network algorithm to correct original infrared image, obtains the image after Nonuniformity Correction and exports.

A neural network nonuniformity means for correcting for short-wave infrared focal plane, is characterized in that, comprising:

Time domain statistics with histogram module, for carrying out time domain statistics with histogram to each pixel in short-wave infrared focal plane;

Neighborhood time domain histogram mean value computation module, for utilize the neighborhood territory pixel of each pixel point time domain histogram calculation described in the equal value histogram of each pixel;

The computing module of neural network estimated value, for being used for the gray scale estimated value of each pixel of neural network algorithm according to the average histogram calculation of described each pixel;

Corrected neural network module, for utilizing neural network algorithm to correct original infrared image, obtaining the image after Nonuniformity Correction and exporting.

The neural network nonuniformity correction method of short-wave infrared focal plane of the present invention and device, by the histogrammic average of neighborhood of pixels statistics time domain, then use the pixel grey scale after the equilibrium of histogram average as the estimated value of neural network algorithm, neural network is finally utilized to complete correction, improve traditional neural network correction method to correct and need the shortcoming of image scene relative motion, and algorithm is simple, computation complexity is low is convenient to engineering practice.

Accompanying drawing explanation

Fig. 1 is the process flow diagram schematic diagram of the neural network nonuniformity correction method of the short-wave infrared focal plane of the embodiment of the present invention.

Fig. 2 is the structural representation of the neural network nonuniformity means for correcting of the short-wave infrared focal plane of the embodiment of the present invention.

Embodiment

Below in conjunction with accompanying drawing, the embodiment of the present invention is described in detail.

Embodiment one

As shown in Figure 1, it is the process flow diagram schematic diagram of the neural network nonuniformity correction method of the short-wave infrared focal plane of the embodiment of the present invention, and the method comprises:

Time domain statistics with histogram step 101: time domain statistics with histogram is carried out to each pixel in short-wave infrared focal plane.

Neighborhood time domain histogram mean value computation step 102: the equal value histogram utilizing each pixel of time domain histogram calculation of the neighborhood territory pixel of each pixel point.Between the pixel supposing short-wave infrared focal plane imaging, gray scale is continuous print, and the difference so in infrared image between neighbor is very little in the statistical significance of certain frame number, this means that two neighbors are almost equal on the histogram of time.According to this hypothesis, we the intermediate value of the response density function regulation of single probe unit to neighborhood territory pixel point response density function, thus realize nonuniformity correction.Definition focal plane pixel (i, j) average response density function ρ _ijy () is the mean value of neighborhood territory pixel point response density function, m, n represent 3 × 3 topography's coordinates, and its formula is as follows:

${\mathrm{\ρ}}_{\mathrm{ij}}\left(y\right)=\frac{1}{8}\underset{m=0}{\overset{2}{\mathrm{\Σ}}}\underset{n=0}{\overset{2}{\mathrm{\Σ}}}{\mathrm{\ρ}}_{\mathrm{ij}}^{\mathrm{mn}}\left(y\right)$ Formula (1)

And average response density function can correspond to the histogram h that all pixel responses are formed _ij(y).Therefore, further the equal value histogram of each for correspondence pixel is designated as h _ijy (), in this step, specifically can utilize the equal value histogram of each pixel of following formulae discovery:

$h_{\mathrm{ij}}\left(y\right)=\frac{1}{8}\underset{m=0}{\overset{2}{\mathrm{\Σ}}}\underset{n=0}{\overset{2}{\mathrm{\Σ}}}{h}_{\mathrm{ij}}^{\mathrm{mn}}\left(y\right)$ Formula (2)

Wherein, the absolute coordinates that (i, j) is pixel, h _ij(y) for absolute coordinates on short-wave infrared focal plane be the equal value histogram of the pixel of (i, j), be the equal value histogram of the pixel neighborhood of a point of (i, j) for absolute coordinates, (m, n) is the relative coordinate of 3 × 3 topographies, wherein, initial value be the time domain histogram of corresponding pixel points.

The calculation procedure 103 of neural network estimated value: the gray scale estimated value being used for each pixel of neural network algorithm according to the average histogram calculation of each pixel.In this step, using equalization histogram as the new histogram of each pixel (i, j), namely inverse transformation completes 3 × 3 topography's time domain equalizations, utilizes this value as the estimated value T of neural network _ij(n).Particularly, in this step, the gray scale estimated value of each pixel of following formulae discovery can specifically be utilized:

$T_{\mathrm{ij}}\left(n\right)=\frac{65535\×\underset{y=0}{\overset{Y_{\mathrm{ij}}\left(n\right)}{\mathrm{\Σ}}}{h}_{\mathrm{ij}}\left(y\right)}{\underset{y=0}{\overset{65535}{\mathrm{\Σ}}}{h}_{\mathrm{ij}}\left(y\right)}$ Formula (3)

Wherein, T _ij(n) for absolute coordinates be the gray scale estimated value of the pixel of (i, j), Y _ij(n) for absolute coordinates in original infrared image be the gray-scale value of the pixel of (i, j), n is the frame number of image.

Corrected neural network step 104: utilize neural network algorithm to correct original infrared image, obtains the image after Nonuniformity Correction and exports.Periodically neural network structure is adopted to carry out the Nonuniformity Correction of single scale, its specific practice allows each neuron connect a pixel, bamboo product hidden layer, its each neuron couples together with contiguous several pixels just as horizontal thin cell element, obtain their average output value and the lower floor's neuron delivering to it goes to calculate heterogeneity, adopt steepest descent method according to real image iteration frame by frame, until reach best correcting state.Particularly, in this step, the image after following formulae discovery Nonuniformity Correction can be utilized to export:

X _ij(n)=w _ij(n) Y _ij(n)+b _ij(n) ... formula (4)

Wherein, w _ijn () is gain correction coefficient, b _ijn () is bias correction coefficient, Y _ijn () is actual measurement gray-scale value, X _ijn () is the gray-scale value after correction, w _ij(n) and b _ijn () is by following equations:

W _ij(n)=w _ij(n-1)-uE _ij(n-1) Y _ij(n-1) ... formula (5)

B _ij(n)=b _ij(n-1)-uE _ij(n-1) ... formula (6)

E _ij(n-1)=T _ij(n-1)-X _ij(n-1) ... formula (7)

Wherein, as n=1, w _ij(n-1), b _ij(n-1), T _ij(n-1), X _ij(n-1) value is designated value, and u is convergence coefficient.

The neural network nonuniformity correction method of the present embodiment short-wave infrared focal plane, by the histogrammic average of neighborhood of pixels statistics time domain, then use the pixel grey scale after the equilibrium of histogram average as the estimated value of neural network algorithm, neural network is finally utilized to complete correction, improve traditional neural network correction method to correct and need the shortcoming of image scene relative motion, and algorithm is simple, computation complexity is low is convenient to engineering practice.

Embodiment two

The present embodiment relates to the neural network nonuniformity means for correcting of the short-wave infrared focal plane corresponding with above-described embodiment one, as shown in Figure 2, this device comprises time domain statistics with histogram module 11, neighborhood time domain histogram mean value computation module 12, the computing module 13 of neural network estimated value and corrected neural network module 14, and the function of modules is as follows:

Time domain statistics with histogram module 11, for carrying out time domain statistics with histogram to each pixel in short-wave infrared focal plane.

Neighborhood time domain histogram mean value computation module 12, for utilizing the equal value histogram of each pixel of time domain histogram calculation of the neighborhood territory pixel of each pixel point.In this module, above-mentioned formula (2) can be utilized to calculate the equal value histogram of each pixel.

The computing module 13 of neural network estimated value, for being used for the gray scale estimated value of each pixel of neural network algorithm according to the average histogram calculation of each pixel.In this module, formula (3) can be utilized to calculate the gray scale estimated value of each pixel.

Corrected neural network module 14, for utilizing neural network algorithm to correct original infrared image, obtaining the image after Nonuniformity Correction and exporting.In this module, above-mentioned formula (4) to formula (7) can be utilized to calculate the image after Nonuniformity Correction and to export

The neural network nonuniformity means for correcting of the present embodiment short-wave infrared focal plane, by the histogrammic average of neighborhood of pixels statistics time domain, then use the pixel grey scale after the equilibrium of histogram average as the estimated value of neural network algorithm, neural network is finally utilized to complete correction, improve traditional neural network correction method to correct and need the shortcoming of image scene relative motion, and algorithm is simple, computation complexity is low is convenient to engineering practice.

The above; be only the specific embodiment of the present invention, but protection scope of the present invention is not limited thereto, is anyly familiar with those skilled in the art in the technical scope that the present invention discloses; change can be expected easily or replace, all should be encompassed within protection scope of the present invention.Therefore, protection scope of the present invention should be as the criterion with the protection domain of described claim.

Claims (8)

1. a neural network nonuniformity correction method for short-wave infrared focal plane, is characterized in that, comprising:

Time domain statistics with histogram step: time domain statistics with histogram is carried out to each pixel in short-wave infrared focal plane;

Neighborhood time domain histogram mean value computation step: the equal value histogram of each pixel described in the time domain histogram calculation utilizing the neighborhood territory pixel of each pixel point;

The calculation procedure of neural network estimated value: the gray scale estimated value being used for each pixel of neural network algorithm according to the average histogram calculation of described each pixel;

Corrected neural network step: utilize neural network algorithm to correct original infrared image, obtains the image after Nonuniformity Correction and exports.

2. method according to claim 1, is characterized in that, in described neighborhood time domain histogram mean value computation step, utilizes the equal value histogram of each pixel described in following formulae discovery:

$h_{\mathrm{ij}}\left(y\right)=\frac{1}{8}\underset{m=0}{\overset{2}{\mathrm{\Σ}}}\underset{n=0}{\overset{2}{\mathrm{\Σ}}}{h}_{\mathrm{ij}}^{\mathrm{mn}}\left(y\right),$

Wherein, the absolute coordinates that (i, j) is pixel, h _ij(y) for absolute coordinates on described short-wave infrared focal plane be the equal value histogram of the pixel of (i, j), be the equal value histogram of the pixel neighborhood of a point of (i, j) for absolute coordinates, (m, n) is the relative coordinate of 3 × 3 topographies, wherein, initial value be the time domain histogram of corresponding pixel points.

3. method according to claim 2, is characterized in that, in the calculation procedure of described neural network estimated value, utilizes the gray scale estimated value of each pixel of following formulae discovery:

$T_{\mathrm{ij}}\left(n\right)=\frac{\mathrm{65535\×}\underset{y=0}{\overset{Y_{\mathrm{ij}}\left(n\right)}{\mathrm{\Σ}}}{h}_{\mathrm{ij}}\left(y\right)}{\underset{y=0}{\overset{65535}{\mathrm{\Σ}}}{h}_{\mathrm{ij}}\left(y\right)},$

Wherein, T _ij(n) for absolute coordinates be the gray scale estimated value of the pixel of (i, j), Y _ij(n) for absolute coordinates in original infrared image be the gray-scale value of the pixel of (i, j), n is the frame number of image.

4. method according to claim 3, is characterized in that, in described corrected neural network step, utilizes the image after following formulae discovery Nonuniformity Correction to export:

X _ij(n)＝w _ij(n)Y _ij(n)+b _ij(n)，

Wherein, w _ijn () is gain correction coefficient, b _ijn () is bias correction coefficient, Y _ijn () is actual measurement gray-scale value, X _ijn () is the gray-scale value after correction, w _ij(n) and b _ijn () is by following equations:

w _ij(n)＝w _ij(n-1)-uE _ij(n-1)Y _ij(n-1)，

b _ij(n)＝b _ij(n-1)-uE _ij(n-1)，

E _ij(n-1)＝T _ij(n-1)-X _ij(n-1)，

Wherein, as n=1, w _ij(n-1), b _ij(n-1), T _ij(n-1), X _ij(n-1) value is designated value, and u is convergence coefficient.

5. a neural network nonuniformity means for correcting for short-wave infrared focal plane, is characterized in that, comprising:

Time domain statistics with histogram module, for carrying out time domain statistics with histogram to each pixel in short-wave infrared focal plane;

Neighborhood time domain histogram mean value computation module, for utilize the neighborhood territory pixel of each pixel point time domain histogram calculation described in the equal value histogram of each pixel;

The computing module of neural network estimated value, for being used for the gray scale estimated value of each pixel of neural network algorithm according to the average histogram calculation of described each pixel;

Corrected neural network module, for utilizing neural network algorithm to correct original infrared image, obtaining the image after Nonuniformity Correction and exporting.

6. device according to claim 5, is characterized in that, in described neighborhood time domain histogram mean value computation module, utilizes the equal value histogram of each pixel described in following formulae discovery:

$h_{\mathrm{ij}}\left(y\right)=\frac{1}{8}\underset{m=0}{\overset{2}{\mathrm{\Σ}}}\underset{n=0}{\overset{2}{\mathrm{\Σ}}}{h}_{\mathrm{ij}}^{\mathrm{mn}}\left(y\right),$

Wherein, the absolute coordinates that (i, j) is pixel, h _ij(y) for absolute coordinates on described short-wave infrared focal plane be the equal value histogram of the pixel of (i, j), be the equal value histogram of the pixel neighborhood of a point of (i, j) for absolute coordinates, (m, n) is the relative coordinate of 3 × 3 topographies, wherein, initial value be the time domain histogram of corresponding pixel points.

7. device according to claim 6, is characterized in that, in the computing module of described neural network estimated value, utilizes the gray scale estimated value of each pixel of following formulae discovery:

$T_{\mathrm{ij}}\left(n\right)=\frac{\mathrm{65535\×}\underset{y=0}{\overset{Y_{\mathrm{ij}}\left(n\right)}{\mathrm{\Σ}}}{h}_{\mathrm{ij}}\left(y\right)}{\underset{y=0}{\overset{65535}{\mathrm{\Σ}}}{h}_{\mathrm{ij}}\left(y\right)},$

Wherein, T _ij(n) for absolute coordinates be the gray scale estimated value of the pixel of (i, j), Y _ij(n) for absolute coordinates in original infrared image be the gray-scale value of the pixel of (i, j), n is the frame number of image.

8. device according to claim 7, is characterized in that, in described corrected neural network step, utilizes the image after following formulae discovery Nonuniformity Correction to export:

X _ij(n)＝w _ij(n)Y _ij(n)+b _ij(n)，

Wherein, w _ijn () is gain correction coefficient, b _ijn () is bias correction coefficient, Y _ijn () is actual measurement gray-scale value, X _ijn () is the gray-scale value after correction, w _ij(n) and b _ijn () is by following equations:

w _ij(n)＝w _ij(n-1)-uE _ij(n-1)Y _ij(n-1)，

b _ij(n)＝b _ij(n-1)-uE _ij(n-1)，

E _ij(n-1)＝T _ij(n-1)-X _ij(n-1)，

Wherein, as n=1, W _ij(n-1), b _ij(n-1), T _ij(n-1), X _ij(n-1) value is designated value, and u is convergence coefficient.