CN100458380C - Non-refrigeration infrared focus plane non-uniform correcting algorithm basedon Wiener filter theory - Google Patents

Abstract

The disclosed correction algorithm based on Wiener filtering theory from in-depth analysis of random signal process theory. Under support of modern signal process technique (mainly, adaptive filtering technique), a structure of digital filter is constructed. The invention applies analysis of filter structure to issue of non-uniform correction of not refrigerative infrared focal plane. Advantages are: higher correction accuracy in real-time.

Description

Non-refrigeration infrared focus plane non-uniform correcting algorithm based on the Wiener filtering theory

Technical field

The present invention relates to the alignment technique of non-refrigerating infrared focal plane imaging system (IRFPA), particularly the nonuniformity correction technology of non-refrigerating infrared focal plane imaging system.

Background technology

For any object in the above temperature of 0K, constitute the molecule of object, atoms etc. all are among the motion of vibration and rotation, and discharge electromagnetic wave to extraneous space.The target of different temperatures, its radiated wave intensity difference.Faint infrared-ray is estimated, and according to the size identification object of infrared-ray intensity, forms image, the principle of non-refrigerating infrared focal plane imaging system that Here it is again.

Pixel how much be a key index of image quality.Obtain the distinct image effect, just need be by the first imaging array that constitutes of many detections.Because technology, each surveys the infrared radiation of unit for same intensity, and the result is inequality in its output, or even difference is very big.Non-refrigeration infrared focus plane non-uniform correcting algorithm promptly is the theoretical method that proposes for addressing this problem.

The non-refrigerating infrared focal plane Nonuniformity Correction is one of gordian technique of infrared focal plane imaging system.Though proposed a variety of correcting algorithms now, as peg method, neural network method etc., these methods or too simple, calibration result not ideal enough (as two point calibrations), or too complex, calculated amount is difficult to realize real-time all have weak point greatly.

At first, IRFPA surveys first temperature rise that is subjected to infrared radiation and causes and can regard a kind of stochastic process as.

Because determining the multidimensional distribution function of a certain class process in actual engineering is suitable difficulty, sometimes or even impossible.Therefore common disposal route is to characterize this process with its statistical nature, and numerical characteristic commonly used is average, variance and related function.We use stochastic process X[t] be illustrated in the temperature rise Δ T in this stochastic process, because temperature rise Δ T＞=0, so its average E[Δ T]=m _T＞0, learn according to theory of random processes, should be to digital Signal Processing earlier to its zero-meanization, the method for zero-meanization can be got all square differential, i.e. first order difference to it.

What the autocorrelation function of stochastic process characterized is its different degrees of correlation of signal constantly, because input signal is identical, there is certain correlativity in its constantly different response.So can make up FIR (finite impulse response) transversal filter on this basis it is carried out signal Processing.

The non-refrigerating infrared focal plane Nonuniformity Correction is one of gordian technique of infrared focal plane imaging system.Though a variety of correcting algorithms have been proposed now, they or too simple, calibration result not ideal enough (as two point calibrations), or too complex, calculated amount is big and be difficult to realize real-time, thereby the some shortcomings part is all arranged.The bearing calibration based on the Wiener filtering theory that this paper proposes not only has higher correction accuracy, and possesses good real time performance.

Summary of the invention

The non-refrigeration infrared focus plane non-uniform correcting algorithm that the present invention proposes based on the Wiener filtering theory, derive from in-depth analysis to the random signal treatment theory, under the support of modern signal processing technology (mainly being auto-adaptive filtering technique), a kind of digital filter configuration that makes up, as shown in Figure 1.The analysis of this filter construction is used for the non-refrigeration infrared focus plane non-uniform correcting problem, higher correction accuracy is not only arranged, and possess good real time performance.

The present invention supposes that circuit of focal plane readout possesses, and is packaged in the optical lens, and signal can be changed sense data and preservation by A/D, and sweep speed is per second 30 frames.The step that this algorithm is used for nonuniformity correction is as follows:

1, determines the Wiener filter order

If, u _i=u (i), i=1,2 ..., m represents m data that independent observation obtains of steady discrete time stochastic process, i.e. the different outputs constantly in focal plane.G (u _i) expression u _iProbability density function.If

Expression u _iConditional probability density function, and the given parameter estimation vector that is used for process model building

If m is a model order, then can write out

${\widehat{\mathrm{\θ}}}_m={[{\widehat{\mathrm{\θ}}}_{1m},{\widehat{\mathrm{\θ}}}_{2m},\·\·\·,{\widehat{\mathrm{\θ}}}_{\mathrm{mm}}]}^T$

Thereby we obtain representing some models of vying each other of process interested.Information theory Standard Selection model with Akaike proposes makes:

$\mathrm{AIC}\left(m\right)=-2L\left({\widehat{\mathrm{\θ}}}_m\right)+2m$

Minimize.Function

Be defined as:

$L\left({\widehat{\mathrm{\θ}}}_m\right)=\max \underset{i=1}{\overset{N}{\mathrm{\Σ}}}\ln f_U\left(u_i\right|{\widehat{\mathrm{\θ}}}_m)$

Ln represents natural logarithm in the formula.

The max log likelihood that is called model parameter is estimated.

Select suitable m value to make the value minimum of function AIC (m), this m value promptly can be used as the exponent number of model.

2, black matrix calibration in the operating temperature range

If the target temperature interval of focal plane detection is T ₁～T _m, with [T ₁, T _m] on average get m temperature spot, be respectively T ₁, T ₂, T ₃T _m

Corresponding to each temperature spot with the black matrix of this temperature as infrared origin (can buy) from market.The temperature of regulating infrared origin is T ₁, and before being placed on target undetermined focal plane camera lens, m data (A read in start ₁₁, A ₁₂, A ₁₃A _1m) and preserve.Shutdown allows the focal plane naturally cool to room temperature.

Repeat this step, the temperature of regulating infrared origin is T ₂, m data (A read in start once more ₂₁, A ₂₂, A ₂₃A _2m) and preserve.Shutdown also allows its natural cooling.

Repeat identical step m time altogether, obtain the matrix A of forming by m * m data.

3, the Wiener filter weights determines

By previous hypothesis, the target temperature interval of focal plane detection is T ₁～T _m,, can think that therefore the temperature when the focal plane is T because formation is gray level image ₁And T _mThe time, the gray shade scale of image is 0 and 2 ^k-1 (k is analog-to-digital exponent number).So corresponding to T ₁Temperature, Expected Response are 2 ^k-1, corresponding to T _mTemperature, Expected Response are 0 (resistance are negative temperature coefficient).In the scope of workspace, IRFPA can regard a linear time invariant system as.Therefore, with [T ₁, T _m] being equally divided into m point, this m the corresponding Expected Response of point promptly is:

b _m＝[(2 ^k-1)……，2×(2 ^k-1)/m，(2 ^k-1)/m，0]

So by system of linear equations: A _MmW _m ^T=b _m ^T

Obtain filter weights: W _m ^T=A _Mm ^-1b _m ^T

4, the output of correction data

Will be by the weights W of the resulting wave filter of step 3 _m ^TWith input vector u (m),, obtain output data through system's real-time operation:

d(m)＝u(m)W _m ^T

D (m) is the gray level image data after the correction.

This algorithm is a kind of theory with comparison widespread use, and concrete embodiment is varied.Can design various signal processing systems, use various digital signal processing chips, all can reach the purpose of nonuniformity correction.

Description of drawings:

Fig. 1: the filter construction of intending employing.

In the filter construction that is adopted, u (n-M)～u (n) is carved into n input constantly, W when being n-M respectively _iBe corresponding weights,

Be n filtering output constantly.

Fig. 2: algorithm implementation structure synoptic diagram.

The front end input of this system is the output of circuit of focal plane readout, and the output result of system is used for final image and shows.

Embodiment:

Below to use the mode of special digital digital signal processing chip PDSP16256 as its use of example brief description.

Special DSP chip PDSP16256, its crystal oscillator frequency F is 40MHz, data throughput is 40M2 ^m, (wherein m is a coefficient relevant with exponent number).Coefficient word length 12bit, the long 16bit of Input Data word (corresponding to AD conversion bit number), output data word length 32bit, the coefficient of succeeding in school is stored among the EPROM.Because the relative general dsp of special DSP chip has very high data throughput, so this structure has good real time performance, can satisfy the requirement of practicability.Shown in this part-structure accompanying drawing 2.

Claims (1)

1, a kind of non-refrigeration infrared focus plane non-uniform correcting algorithm based on the Wiener filtering theory is characterized in that may further comprise the steps:

1), determine the Wiener filter order:

If u _i=u (i) i=1,2 ... m, represent m data that independent observation obtains of steady discrete time stochastic process, i.e. the different outputs constantly in focal plane, g (u _i) expression u _iProbability density function,

Expression u _iConditional probability density function, and the given parameter estimation vector that is used for process model building

If m is a model order, then

${\widehat{\mathrm{\θ}}}_m={[{\widehat{\mathrm{\θ}}}_{1m},{\widehat{\mathrm{\θ}}}_{2m},...{\widehat{\mathrm{\θ}}}_{\mathrm{mm}}]}^T$

Thereby we obtain representing some models of vying each other of process interested, and the information theory Standard Selection model with Akaike proposes makes:

$\mathrm{AIC}\left(m\right)=-2L\left({\widehat{\mathrm{\θ}}}_m\right)+2m$

Minimize;

Function

Be defined as:

$L\left({\widehat{\mathrm{\θ}}}_m\right)=\max \underset{i=1}{\overset{N}{\mathrm{\Σ}}}\ln f_U\left(u_i\right|{\widehat{\mathrm{\θ}}}_m)$

Ln represents natural logarithm in the formula,

The max log likelihood that is called model parameter is estimated; Select suitable m value to make the value minimum of function AIC (m), this m value is the exponent number of model;

2), black matrix calibration in the operating temperature range:

If the target temperature interval of focal plane detection is T ₁～T _m, with [T ₁, T _m] on average get m temperature spot, be respectively T ₁, T ₂, T ₃..., T _m

Corresponding to each temperature spot with the black matrix of this temperature as infrared origin, the temperature of regulating infrared origin is T ₁, and before being placed on target undetermined focal plane camera lens, m data A read in start ₁₁, A ₁₂, A ₁₃..., A _1mAnd preserve, shutdown allows the focal plane naturally cool to room temperature;

Repeat this step, the temperature of regulating infrared origin is T ₂, m data A read in start once more ₂₁, A ₂₂, A ₂₃..., A _2mAnd preserve, shut down and allow its natural cooling;

Repeat identical step m time altogether, obtain the matrix A that m * m data are formed;

3), determining of Wiener filter weights:

By previous hypothesis, the target temperature interval of focal plane detection is T ₁～T _m,, think that therefore the temperature when the focal plane is T because formation is gray level image ₁And T _mThe time, the gray shade scale of image is 0 and 2 ^k-1, k is analog-to-digital exponent number, so corresponding to T ₁Temperature, Expected Response are 2 ^k-1, corresponding to T _mTemperature, Expected Response are 0, and resistance is negative temperature coefficient;

In the scope of workspace, regard IRFPA as a linear time invariant system, therefore, with [T ₁, T _m] being equally divided into m point, this m the corresponding Expected Response of point promptly is:

b _m＝[(2 ^k-1)，...，2×(2 ^k-1)/m，(2 ^k-1)/m，0]

So by system of linear equations: $A_{\mathrm{mm}}{W}_m^T={b_m}^T$

Obtain filter weights: $W_m^T={A_{\mathrm{mm}}}^{-1}{b_m}^T$

4), the output of correction data:

Will be by the weights W of the resulting wave filter of step 3) _m ^TWith input vector u (m),, obtain output data through being around real-time operation:

$d\left(m\right)=u\left(m\right)W_m^T$

D (m) is the gray level image data after the correction.

Images