CN102013090A - Passive millimetre wave image strip noise suppression method - Google Patents

Abstract

The invention discloses a passive millimetre wave image strip noise suppression method, wherein a neural network algorithm performance is enhanced obviously via improving an estimated signal true value expression; a gain compensation coefficient vector Gc+1 and an offset compensation coefficient vector Oc+1 calculated by iteration reflect gains and offset differences between different lines well. Compared with other corrosion method, a smaller signal to noise ratio can be allowed, the correction performance is improved greatly, and the strip noise can be suppressed strongly; for the convergence property of algorithm, a plurality of detailed information can be preserved, and images are not distorted for overcorrection, therefore, the method is quite suitable for multi-beam millimetre wave imaging.

Description

A kind of passive millimeter-wave image band noise suppressing method

Technical field

The invention belongs to passive mm-wave imaging technical field, more specifically say, relate to a kind of passive millimeter-wave image band noise suppressing method.

Background technology

Focal plane linear array radiometers image-forming system adopts multiple-beam antenna that object scene is carried out scanning imagery.There is heterogeneity in many receiving cables of multiple-beam antenna correspondence, and promptly under identical emittance initial conditions, each passage output voltage values is also inequality.

Fig. 1 is following 4 the non-homogeneous output waveform figures of passage of identical initial conditions.

As shown in Figure 1, this waveform is the signal output waveform of synchronization 4 passages when importing same triangular wave, and wherein the 1-210 position is the output signal of first passage on the horizontal ordinate, is second passage output signal on the 211-420 position, each signal has 210 sampled points, by that analogy.Ignore The noise, can see, the gain and the biasing of each passage there are differences, i.e. the unevenness of passage.This unevenness shows as the band noise in imaging process.

The passage Nonuniformity Correction can suppress the band noise, is divided into based on the correction of calibration with based on the correction of scene.When design, even used correction way based on calibration, it is in full accord that but the reference source of the correspondence of each passage can not be accomplished, can not proofread and correct the unevenness of eliminating passage fully by calibration, suppresses the band noise effectively, therefore, correction based on calibration is not enough as an external calibration of system imaging, also need add once based on the calibration of scene in the system imaging process again, and calibration promptly decided at the higher level but not officially announced just, further eliminate the passage heterogeneity, suppress the band noise.

At present, be used for mainly containing based on the method for the Nonuniformity Correction of scene: histogram is proofreaied and correct, the square coupling is proofreaied and correct, frequency domain filtering is proofreaied and correct and small echo denoising etc.

1, histogram is proofreaied and correct

Histogram is proofreaied and correct and is supposed that at first the target area that each sensor is surveyed has identical radiation profiles, thus the histogram of the subimage of single-sensor acquisition is adjusted to the reference histograms shape, to reach the purpose of denoising.But in fact the intensity profile of each number of sub images is also different, and histogram matching forces to adjust the histogram distribution of original image, has often destroyed the raw information of image, and denoising effect is relatively poor in actual applications.

2, the square coupling is proofreaied and correct

Square coupling calibration result is better than the histogram matching method, but its uniformity requirement to the gradation of image distribution is also very high.Distribute at system's visual field internal object that complexity causes under the uneven situation of intensity profile, use the square coupling to proofread and correct in the time of can producing usually banded characteristic bright when dark, band noise denoising performance is also undesirable.

3, frequency domain filtering is proofreaied and correct

Frequency domain filtering is proofreaied and correct and is difficult to the selection of filter parameter, and the band noise is complicated in the distribution of frequency domain, and distribution is all arranged from the high frequency to the low frequency part.And frequency domain filtering selects also can to cause losing of the original effective information of image when wrong at filter cutoff frequency, and this usually loses more than gain for the original fewer millimeter-wave image of spectrum information.

4, the small echo denoising is proofreaied and correct

Have the scholar to mention though the small echo denoising is proofreaied and correct, all do not carry out further investigation, because the band noise is not simple additive noise, not only have additive noise, also have multiplicative noise, the use of small echo is restricted.

Sum up above several corrections and suppress band noise method, there is common defective in they is exactly relatively poor to the effect of band Noise Suppression, and the algorithm that uses is dispersed the improper picture quality variation that can make on the contrary of denoising easily.

Summary of the invention

The objective of the invention is to overcome the deficiencies in the prior art, a kind of passive millimeter-wave image band noise suppressing method that suppresses better effects and if can keep a large amount of detailed information is provided.

For achieving the above object, the passive millimeter-wave image band of the present invention noise suppressing method is characterized in that, may further comprise the steps:

(1), for the M that has the band noise capable * the passive millimeter-wave image X of N row, gain compensation coefficient vector G is set _c, bias compensation coefficient vector O _c, wherein, gain compensation coefficient vector G _c, bias compensation coefficient vector O _cBe the column vector of M * 1, subscript c represents iterations; Initialization iterations c=1, gain compensation coefficient vector G _cFor vector of unit length (1,1,1 ... 1) ^T, bias compensation coefficient vector O _cFor null vector (0,0,0 ... 0) ^TInitialization columns k=1, initialization iteration step length λ;

(2), get passive millimeter-wave image X k column data as input signal vector μ _k:

μ _k＝(X(1，k)，X(2，k)，…，X(M，k)) ^T (1)；

(3), utilize gain compensation coefficient vector G _cWith bias compensation coefficient vector O _cCorrection has the input signal vector μ of band noise _k, obtain proofreading and correct output vector

${\widehat{\mathrm{\μ}}}_k\left(i\right)=G_c\left(i\right){\mathrm{\μ}}_k\left(i\right)+O_c\left(i\right)---\left(2\right);$

Wherein, Be k row input signal vector μ _kI element μ _k(i) correction of gray-scale value output, G _c(i) be the c time iteration gain penalty coefficient vector G _cI element, O _c(i) be the c time iteration bias compensation coefficient vector O _cI element;

(4), the correction output vector of utilizing step (3) to obtain

The estimated signal true value

${\hat{f}}_k\left(i\right)=({\widehat{\mathrm{\μ}}}_k(i-1)+{\widehat{\mathrm{\μ}}}_k(i+1))/2---\left(3\right);$

(5), calculate the gain compensation coefficient vector G that upgrades _C+1, bias compensation coefficient vector O _C+1:

$G_{c+1}\left(i\right)=G_c\left(i\right)-2\mathrm{\λ}{\mathrm{\μ}}_k\left(i\right)({\widehat{\mathrm{\μ}}}_k\left(i\right)-{\hat{f}}_k\left(i\right))$

$O_{c+1}\left(i\right)=O_c\left(i\right)-2\mathrm{\λ}({\widehat{\mathrm{\μ}}}_k\left(i\right)-{\hat{f}}_k\left(i\right))$ (4)；

(6) if iterations c=C _Max, then stop iteration, carry out step (7), otherwise, make k=k+1, c=c+1, G _c(i)=G _C+1(i), O _c(i)=O _C+1(i), return step (2), wherein, C _MaxBe integer less than passive millimeter-wave image X columns N;

(7) with the last gain compensation coefficient vector G that obtains _C+1, bias compensation coefficient vector O _C+1Correction exist the M of band noise capable * the passive millimeter-wave image X of N row, the passive millimeter-wave image Y of the band squelch after obtaining proofreading and correct

Y(i，j)＝G _c+1(i)X(i，j)+O _c+1(i) (5)

Wherein, (i j) is the gray-scale value of the capable j column element of passive millimeter-wave image Y i of the band squelch after proofreading and correct to Y, and (i j) is the gray-scale value that has the capable j column element of passive millimeter-wave image X i of band noise, G to X _C+1(i) be gain compensation coefficient vector G after c iteration _C+1I element value, O _C+1(i) be biasing penalty coefficient vector O after c iteration _C+1I element value, i=1,2,3 ..., M; J=1,2,3 ... N.

The object of the present invention is achieved like this:

Passive millimeter-wave image hyperchannel heterogeneity mathematical model, as shown in Equation (6):

X(i，j)＝a _iI(i，j)+b _i (6)

Wherein I is the passive millimeter-wave image of desirable no band noise, and the image size is M * N, and M is a picturedeep, number of channels just, and N is a picturewide, the number of scan points when also being horizontal scanning.a _iBe the gain coefficient of passage i, b _iBe passage i biasing coefficient.The i of the output correspondence image of passage i is capable.a _i(i=1,2 ... M) constitute gain coefficient vector α=(a ₁, a ₂..., a _M), and establish Normal Distribution N (1,0.1); b _i(i=1 2...M) constitutes biasing coefficient vector β=(b ₁, b ₂... b _M), and Normal Distribution N (1,0.2).X exists the M of band noise capable * the passive millimeter-wave image of N row, X (i, j) be the gray-scale value that has the capable j column element of passive millimeter-wave image X i of band noise, because system's along continuous straight runs scanning imagery, the i of the signal correspondence image that a certain passage i scanning obtains is capable, gain and bigotedly can regard constant as only exists gain and biasing difference, so the distribution in horizontal direction on obtaining image of band noise between different rows.

Therefore, the present invention is used for passive millimeter-wave image band noise processed with neural network algorithm, by improving estimated signal true value expression formula, the neural network algorithm performance is significantly improved, the gain compensation coefficient vector G that iterative computation goes out _C+1, bias compensation coefficient vector O _C+1The gain and the biasing difference that exist between the different rows have been reacted well.Than other bearing calibrations, can allow littler signal to noise ratio (S/N ratio), and correcting feature is largely increased, the band noise can access stronger inhibition; Because the convergence of algorithm characteristic can keep a large amount of detailed information, can not cause image distortion because of excessive correction, be fit to very much the multi-beam mm-wave imaging and use.

Description of drawings

Fig. 1 is following 4 the non-homogeneous output waveform figures of passage of identical initial conditions;

Fig. 2 is the process flow diagram of the passive millimeter-wave image band of the present invention noise suppressing method;

Fig. 3 is the effect comparison figure of the various band noise suppressing methods under the embodiment of the invention;

Fig. 4 is the effect comparison figure of the various band noise suppressing methods under another embodiment of the present invention;

Fig. 5 is the graph of a relation of iterations and improvement factor Q under the passive millimeter-wave image band of the present invention noise suppressing method one embodiment.

Embodiment

Below in conjunction with accompanying drawing the specific embodiment of the present invention is described, so that those skilled in the art understands the present invention better.What need point out especially is that in the following description, when perhaps the detailed description of known function and design can desalinate main contents of the present invention, these were described in here and will be left in the basket.

Embodiment

Fig. 2 is the process flow diagram of the passive millimeter-wave image band of the present invention noise suppressing method.

As shown in Figure 1, in the present embodiment, its flow process is consistent with the described step of summary of the invention, does not repeat them here.Need to prove that columns k initialization value can be any integer less than image X columns N, its renewal in each iteration also can be calculated according to other modes, as the mode etc. of successively decreasing.

In the present embodiment, the M that has the band noise that need to proofread and correct is capable * simple scenario image that the passive millimeter-wave image X of N row is 256 * 256 sizes, initialization iteration step length λ is 0.1.

Example 1

Fig. 3 is the effect comparison figure of the various band noise suppressing methods under the embodiment of the invention.In this example, at be simple scenario, wherein, Fig. 3 (a) does not have the passive millimeter-wave image of band noise for simple scenario, Fig. 3 (b) is for adding the passive millimeter-wave image of band noise, and Fig. 3 (c) is the result of square coupling band noise suppressing method, and Fig. 3 (d) is the result of frequency domain filtering band noise suppressing method, Fig. 3 (e) is the result of small echo denoising band noise suppressing method, and Fig. 3 (f) is the result of band noise suppressing method of the present invention.

Superiority for the bearing calibration of verifying and illustrate the relative prior art of the present invention, in this example, the band noise that brings for analog channel is inhomogeneous, the average that each provisional capital on the passive millimeter-wave image of the no band noise shown in Fig. 3 (a) be multiply by Gaussian distributed is 1, variance is 0.01 gain coefficient, the average that also adds Gaussian distributed is 0, and variance is 0.01 biasing coefficient, simulates the passive millimeter-wave image of the adding band noise shown in Fig. 3 (b).

By observing Fig. 3 (c), (d), (e), (f), we can see, frequency domain filtering band noise suppressing method and band noise suppressing method of the present invention, the band noise suppression effect is comparatively desirable, small echo denoising band noise suppressing method has been eliminated the band noise of part, but still has significantly horizontal stripe in the passive millimeter-wave image.

For the treatment effect of the various band noise suppressing methods of quantitative evaluation, defined an image enhancement factor Q, suc as formula shown in, Q is big more, it is good more to improve effect, the approaching more original passive millimeter-wave image of the image after the band squelch.

$Q\stackrel{\mathrm{\Δ}}{=}10{\log }_{10}\left(\frac{\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}{d_s}^2[i,j]}{\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}{d_e}^2[i,j]}\right)---\left(1\right)$

d _s[i，j]＝X[i，j]-I[i，j]

d _e[i，j]＝Y[i，j]-I[i，j]

At the image of this simple scenario of pistol image, the image enhancement factor of several band noise suppressing methods is:

Example 2

Fig. 4 is the effect comparison figure of the various band noise suppressing methods under another embodiment of the present invention.In this example, at be the complex scene image, wherein, Fig. 4 (a) does not have the passive millimeter-wave image of band noise for complex scene, Fig. 4 (b) is for adding the passive millimeter-wave image of band noise, and Fig. 4 (c) is the result of square coupling band noise suppressing method, and Fig. 4 (d) is the result of frequency domain filtering band noise suppressing method, Fig. 4 (e) is the result of small echo denoising band noise suppressing method, and Fig. 4 (f) is the result of band noise suppressing method of the present invention.

In order to verify the good characteristic that can keep image detail information in the passive millimeter-wave image band of the present invention noise suppressing method noise suppression process, the passive millimeter-wave image of no band noise is the complex scene image of 256 * 256 sizes, shown in Fig. 4 (a), this image is the part of human body, and this people covers at clothes and hides leader's rifle and a slice derby down.

The average that each provisional capital on the passive millimeter-wave image of the no band noise shown in Fig. 4 (a) be multiply by Gaussian distributed is 1, variance is 0.01 gain coefficient, the average that also adds Gaussian distributed is 0, variance is 0.01 biasing coefficient, simulates the passive millimeter-wave image of the adding band noise shown in Fig. 4 (b).

Can see that by observing we can see that square coupling band noise suppressing method is owing to fail to satisfy applicable elements, filtering band noise on request; Small echo denoising band noise suppressing method has been eliminated the band noise of part, but still has significantly horizontal stripe in the image; The inhibition effect of frequency domain filtering band noise suppressing method and band noise suppressing method of the present invention is comparatively obvious, but careful contrast, can see, frequency domain filtering band noise suppressing method makes passive millimeter-wave image lose detailed information, blured such as details figures such as bracelets on the fold on the clothes, the wrist, and band noise suppressing method of the present invention has not only effectively suppressed the band noise, makes these image details obtain keeping simultaneously, and effect is very desirable.

Conceal the image of this complex scene of pistol image at human body, the image enhancement factor of several band noise denoise algorithm is:

Fig. 5 is the graph of a relation of iterations and improvement factor Q under the passive millimeter-wave image band of the present invention noise suppressing method one embodiment.

Among Fig. 5, the corresponding improvement factor Q of each iteration draws iterations-improvement factor curve then, obtains the graph of a relation of iterations and improvement factor Q.In Fig. 5, we as can be seen, the passive millimeter-wave image band of the present invention noise suppressing method is along with the increase of iterations, the image enhancement factor is dull increase trend, and finally converge on a constant, therefore, the passive millimeter-wave image band of this present invention noise suppressing method is a convergent.

Although above the illustrative embodiment of the present invention is described; so that the technician of present technique neck understands the present invention; but should be clear; the invention is not restricted to the scope of embodiment; to those skilled in the art; as long as various variations appended claim limit and the spirit and scope of the present invention determined in, these variations are conspicuous, all utilize innovation and creation that the present invention conceives all at the row of protection.

Claims (2)

1. a passive millimeter-wave image band noise suppressing method is characterized in that, may further comprise the steps:

(1), for the M that has the band noise capable * the passive millimeter-wave image X of N row, gain compensation coefficient vector G is set _c, bias compensation coefficient vector O _c, wherein, gain compensation coefficient vector G _c, bias compensation coefficient vector O _cBe the column vector of M * 1, subscript c represents iterations; Initialization iterations c=1, gain compensation coefficient vector G _cFor vector of unit length (1,1,1 ... 1) ^T, bias compensation coefficient vector O _cFor null vector (0,0,0 ... 0) ^TInitialization columns k=1, initialization iteration step length λ;

(2), get passive millimeter-wave image X k column data as input signal vector μ _k:

μ _k＝(X(1，k)，X(2，k)，…，X(M，k)) ^T (1)；

(3), utilize gain compensation coefficient vector G _cWith bias compensation coefficient vector O _cCorrection has the input signal vector μ of band noise _k, obtain proofreading and correct output vector

${\widehat{\mathrm{\μ}}}_k\left(i\right)=G_c\left(i\right){\mathrm{\μ}}_k\left(i\right)+O_c\left(i\right)---\left(2\right);$

Wherein,

Be k row input signal vector μ _kI element μ _k(i) correction of gray-scale value output, G _c(i) be the c time iteration gain penalty coefficient vector G _cI element, O _c(i) be the c time iteration bias compensation coefficient vector O _cI element;

(4), the correction output vector of utilizing step (3) to obtain

, the estimated signal true value

${\hat{f}}_k\left(i\right)=({\widehat{\mathrm{\μ}}}_k(i-1)+{\widehat{\mathrm{\μ}}}_k(i+1))/2---\left(3\right);$

(5), calculate the gain compensation coefficient vector G that upgrades _C+1, bias compensation coefficient vector O _C+1:

$G_{c+1}\left(i\right)=G_c\left(i\right)-2\mathrm{\λ}{\mathrm{\μ}}_k\left(i\right)({\widehat{\mathrm{\μ}}}_k\left(i\right)-{\hat{f}}_k\left(i\right))$

$O_{c+1}\left(i\right)=O_c\left(i\right)-2\mathrm{\λ}({\widehat{\mathrm{\μ}}}_k\left(i\right)-{\hat{f}}_k\left(i\right))$ (4)；

(6) if iterations c=C _Max, then stop iteration, carry out step (7), otherwise, make k=k+1, c=c+1, G _c(i)=G _C+1(i), O _c(i)=O _C+1(i), return step (2), wherein, C _MaxBe integer less than passive millimeter-wave image X columns N;

(7) with the last gain compensation coefficient vector G that obtains _C+1, bias compensation coefficient vector O _C+1Correction exist the M of band noise capable * the passive millimeter-wave image X of N row, the passive millimeter-wave image Y of the band squelch after obtaining proofreading and correct

Y(i，j)＝G _c+1(i)X(i，j)+O _c+1(i) (5)

Wherein, (i j) is the gray-scale value of the capable j column element of passive millimeter-wave image Y i of the band squelch after proofreading and correct to Y, and (i j) is the gray-scale value that has the capable j column element of passive millimeter-wave image X i of band noise, G to X _C+1(i) be gain compensation coefficient vector G after c iteration _C+1I element value, O _C+1(i) be biasing penalty coefficient vector O after c iteration _C+1I element value, i=1,2,3 ..., M; J=1,2,3 ... N.

2. passive millimeter-wave image band noise suppressing method according to claim 1 is characterized in that described iteration step length λ initialization value is 0.1.

Images