CN101582159A - Infrared image background suppression method based on unsupervised kernel regression analysis - Google Patents

Abstract

The invention discloses an infrared image background suppression method based on unsupervised kernel regression analysis, belonging to the image processing field. The infrared image background suppression method solves the technical problems that priori knowledge is needed and self-adaptability is poor in the field of infrared image background clutter suppression. Firstly, a sliding window is set to predict background, and Gaussian function is determined to serve as kernel function for unsupervised kernel regression analysis; a background predicting clutter sample is substituted in the function to calculate unsupervised kernel regression equation, and a central test sample (O) is input into the unsupervised kernel regression equation so as to obtain the predicted value of the central test sample (O); then, the central test sample (O) value subtracts the predicted value; the sliding window moves, the above process is repeated until the whole image is processed, and a background suppression result image is output. The invention can effectively improve target detectability and positioning accuracy of an infrared target recognition and tracking system, an infrared image monitoring system, etc.

Description

Based on the infrared image background inhibition method of not having the supervision kernel regression analysis

Technical field

The present invention relates to a kind of infrared image background inhibition method, belong to image processing field.

Background technology

In infrared automatic target designation system, in order to find target as early as possible, make infrared guidance system that the enough reaction time be arranged and improve the early warning distance of the weapons of system, require on far distance, just can detect target, thereby can find target as early as possible.When detection range and imaging viewing field increase,, in imaging plane, also only show as several pixels, even, be called little target less than a pixel even target itself is very big.At this moment, detectable signal relatively a little less than, particularly under the varying background of non-stationary is disturbed, target even flooded by the noise of large amount of complex (clutter), signal noise ratio (snr) of image is extremely low, makes Point Target Detection work become very difficult.Therefore, for outstanding little target, improve signal to noise ratio (S/N ratio), thereby improve target detection probability, the background before the infrared small target image is detected suppresses and noise filtering is very necessary.In the automatic target recognition system, background suppresses and noise-cut is referred to as the preceding filter preprocessing of detection.The final purpose that infrared image background suppresses is in order to eliminate noise jamming, to preserve target information.

At present, in infrared image background suppressed to handle, airspace filter was outbalance, widely used one big class methods.Typical airspace filter method comprises linear background forecast, Top-hat conversion (morphologic filtering), nonlinear filtering (mean filter, gaussian filtering) etc., they realize the prediction of infrared image background by airspace filter, utilize prognostic chart and original graph to do difference and obtain the result that background suppresses, thereby reach the purpose that suppresses background.

In recent years, neural network, support vector regression etc. has the supervision machine learning method also to be applied to gradually in the infrared image background inhibition processing, and has obtained and suppressed effect preferably.But the supervised learning method needs a large amount of prioris (being training sample), carries out the learning training that background suppresses model in advance, can't satisfy no priori situation in the actual treatment, and not have adaptivity.

Summary of the invention

The present invention for solve in infrared image background clutter inhibition field, exist need priori, the relatively poor technical barrier of adaptivity, provide a kind of based on the infrared image background inhibition method of not having the supervision kernel regression analysis.

The present invention includes following steps:

Step 1, setting moving window are used for background forecast; Moving window adopts double window mouth pattern, and the center of interior window is the center test sample book; Sample in the exterior window is a projected background clutter sample;

Step 2, employing Gaussian function are as the kernel function of not having the supervision kernel regression analysis;

Step 3, the projected background clutter sample information of utilizing current exterior window are updated in the kernel function of not having the supervision kernel regression analysis as the regression data sample value, and calculating does not have supervision nuclear regression equation;

Step 4, the center test sample book information of window in current is input to described nothing supervision nuclear regression equation, obtains the prediction clutter gray-scale value of center test sample book;

Step 5, utilize current in the center test sample book gray-scale value of window deduct the prediction clutter gray-scale value of the center test sample book that step 4 obtains, thereby suppress the background clutter of infrared image;

Step 6, mobile moving window, its moving step length is 1, returns step 3, up to the traversal full figure, the infrared image after the output background suppresses.

The present invention compared with prior art has following advantage:

(1) adopts the nuclear regression technique, can handle strong fluctuating, complicated infrared background clutter data effectively, have good nonlinear data predictive ability.

(2) need not precondition regression model (promptly need not the training sample priori), carry out unsupervised learning, have good local auto-adaptive predictive ability according to the test pattern its data.

(3) adopt the double window mouth that the infrared image regional area is carried out regression forecasting, it is good to suppress background effect, effectively preserves important goal simultaneously, can improve target detection ability and bearing accuracy as systems such as infrared identification and tracking, infrared image monitorings greatly.

Description of drawings

Fig. 1 is the synoptic diagram of moving window; Fig. 2 is the position view that the center test sample book O of the interior window B of moving window is positioned at the edge of infrared image C.

Embodiment

Embodiment one: in conjunction with Fig. 1 present embodiment is described, the present embodiment step is as follows:

Step 1, setting moving window are used for background forecast; Moving window adopts double window mouth pattern background clutter to be predicted moving window is made up of interior window B and exterior window A; Interior window B is positioned at the center of moving window, the center of interior window B is center test sample book O, interior window B is used to protect the center test sample book O information that is positioned at the moving window center, interior window B zone beyond the test sample book O of center is equivalent to the protection zone, prevent at test sample book O to be to select under the situation of object pixel to choose the sample relevant in the process of background clutter sample, so produced the protection zone with target; The outside of window B is exterior window A in the moving window, and the sample among the exterior window A is a projected background clutter sample, and exterior window A is used to select projected background clutter sample information;

Step 2, employing Gaussian function are as the kernel function of not having the supervision kernel regression analysis; Be used to measure the similarity between the background clutter sample;

Step 3, the projected background clutter sample information of utilizing current exterior window A are updated in the kernel function of not having the supervision kernel regression analysis as the regression data sample value, and calculating does not have supervision nuclear regression equation;

Step 4, the center test sample book O information of window B in current is input to described nothing supervision nuclear regression equation, obtains the prediction clutter gray-scale value of center test sample book O;

Step 5, utilize current in the gray-scale value of center test sample book O of window B deduct the prediction clutter gray-scale value of the center test sample book O that step 4 obtains, thereby suppress the background clutter of infrared image; If the center test sample book O of window B is a background clutter in current, the gray-scale value of then described center test sample book O will be removed; If the center test sample book O of window B is an object pixel in current, the background clutter gray-scale value in then suppressing to be superimposed upon on the center test sample book O of window B;

Step 6, mobile moving window, its moving step length is 1, returns step 3, up to the traversal full figure, the infrared image after the output background suppresses.

Embodiment two: present embodiment is described in conjunction with Fig. 2, present embodiment and embodiment one difference are if work as the edge that the center test sample book O of the interior window B of pre-treatment is positioned at infrared image C, then the sample of the part of the disappearance in the moving window adopts the mirror image symmetric mode to obtain, and promptly adopts the mirror image symmetric mode to obtain to lack the gray-scale value of partial pixel point.Other step is identical with embodiment one.

Embodiment three: it is as follows that present embodiment and embodiment one difference are to try to achieve in the step 3 step of not having supervision nuclear regression equation:

The recurrence estimation formulas is as follows:

y _i＝z(x _i)+ε _i，i＝1，……，P，(1)

Wherein, x _iBe the 2x1 dimensional vector, the coordinate of expression two-dimensional space, y _iRepresent corresponding gradation of image value; Z (x _i) be called regression function, ε _iBe stochastic error or random disturbance, it is one and distributes and x _iIrrelevant stochastic variable, it is that average is 0 normally distributed random variable; With z (x _i) launch at neighborhood, can obtain following formula:

z(x _i)＝β ₀+β ₁ ^T(x _i-x)+β ₂ ^Tvech{(x _i-x)(x _i-x) ^T}+… (2)

Definition vech () handles the vectorization of triangular portions under the symmetric matrix, with 2 * 2 symmetrical matrix:

$\mathrm{vech}\left(\left[\begin{array}{cc}a& b\\ b& d\end{array}\right]\right)={\left[\begin{array}{ccc}a& b& d\end{array}\right]}^T---\left(3\right)$

β ₀=z (x), β ₁And β ₂Satisfy:

${\mathrm{\β}}_1=\▿z\left(x\right)={[\frac{\∂z\left(x\right)}{\∂x_1},\frac{\∂z\left(x\right)}{\∂x_2}]}^T---\left(4\right)$

${\mathrm{\β}}_2=\frac{1}{2}{[\frac{{\∂}^{2}z\left(x\right)}{\∂x_1^2},2\frac{{\∂}^{2}z\left(x\right)}{\∂x_1{x}_2},\frac{{\∂}^{2}z\left(x\right)}{\∂x_2^2}]}^T---\left(5\right)$

And parameter beta _nObtain by finding the solution following optimization problem:

$\underset{\left\{{\mathrm{\β}}_n\right\}}{\min }\underset{i=1}{\overset{P}{\mathrm{\Σ}}}{[y_i-{\mathrm{\β}}_0-{\mathrm{\β}}_1^T(x_i-x)-{\mathrm{\β}}_2^T\mathrm{vech}\{(x_i-x){(x_i-x)}^T\}-\·\·\·]}^2.---\left(6\right)$

$K_H(x_i-x,y_i-y)$

Wherein

$K_H\left(t\right)=\frac{K\left(H^{1}t\right)}{\det \left(H\right)}---\left(7\right)$

K _H(t) be the nuclear weighting function, H is called smoothing matrix;

Utilize the mathematical operation method to carry out abbreviation, the zeroth order estimated value of trying to achieve z (x) is:

$\hat{z}\left(x\right)={\widehat{\mathrm{\β}}}_0=e_1^T{\left(X_x^T{W}_x{X}_x\right)}^{-1}{X}_x^T{W}_{x}y---\left(8\right)$

E wherein ₁ ^TBe that one first row element is 1, other is 0 column vector,

y＝[y ₁，y ₂，…，y _P] ^T (9)

W _x＝diag[K _H(x ₁-x)，K _H(x ₂-x)，…，K _H(x _P-x)] (10)

$X_x=\left[\begin{array}{ccccc}1& {(x_1-x)}^T& {\mathrm{vech}}^T\left\{(x_1-x){(x_1-x)}^T\right\}& \·\·\·& \\ 1& {(x_2-x)}^T& {\mathrm{vech}}^T\left\{(x_2-x){(x_2-x)}^T\right\}& & \·\·\·\\ \·& \·& \·& \·& \\ \·& \·& \·& \·& \\ \·& \·& \·& \·& \\ 1& {(x_P-x)}^T& {\mathrm{vech}}^T\left\{(x_P-x){(x_P-x)}^T\right\}& & \·\·\·\end{array}\right]---\left(11\right)$

By top formula as can be seen, the result of evaluation

Depend in part on the selection of smoothing matrix; Here, use a simple and higher model of counting yield to represent:

H _i＝hu _iI (12)

In following formula, u _iCharacterization data sampling dense degree (generally makes u _i=1), h is called smoothing parameter, and its value is calculated by the series of iterations formula; Easy in order to calculate, generally speaking, the value of h is near " 2 ";

Like this,, and calculate,, can obtain the recurrence estimated value of local neighborhood center pixel result of calculation substitution formula 8 according to formula 9 to formula 12 as long as determine interior each gray values of pixel points of image local neighborhood; Usually, kernel function all adopts the Gaussian radial basis function form; Return kernel function form K and parameter h and u this moment _iAs follows: K is the gaussian kernel function form, $K(x,y)=\exp (-\frac{{\left|\right|x-y\left|\right|}^2}{{\mathrm{\σ}}^2}),$ The h value is 2, u _iValue is 1.The projected background clutter sample that does not have supervision nuclear regression equation and be a current exterior window A as the regression data sample value be updated to formula 3 to formula 7 and formula 9 to formula 12, calculate X _XAnd W _XValue after, be updated in the formula 8 resulting.Other step is identical with embodiment one.

Content of the present invention is not limited only to the content of the respective embodiments described above, and the combination of one of them or several embodiments equally also can realize the purpose of inventing.

Claims (5)

1, based on the infrared image background inhibition method of not having the supervision kernel regression analysis, it is characterized in that its step is as follows:

Step 1, setting moving window are used for background forecast; Moving window adopts double window mouth pattern, and the center of interior window (B) is center test sample book (O); Sample in the exterior window (A) is a projected background clutter sample;

Step 2, employing Gaussian function are as the kernel function of not having the supervision kernel regression analysis;

Step 3, the projected background clutter sample information of utilizing current exterior window (A) are updated in the kernel function of not having the supervision kernel regression analysis as the regression data sample value, and calculating does not have supervision nuclear regression equation;

Step 4, center test sample book (O) information of window (B) in current is input to described nothing supervision nuclear regression equation, obtains the prediction clutter gray-scale value of center test sample book (O);

Step 5, utilize current in center test sample book (O) gray-scale value of window (B) deduct the prediction clutter gray-scale value of the center test sample book (O) that step 4 obtains, thereby suppress the background clutter of infrared image;

Step 6, mobile moving window, its moving step length is 1, returns step 3, up to the traversal full figure, the infrared image after the output background suppresses.

2, according to claim 1 based on the infrared image background inhibition method of not having the supervision kernel regression analysis, it is characterized in that then the sample of the part of the disappearance in the moving window adopts the mirror image symmetric mode to obtain if work as the edge that the center test sample book (O) of the interior window (B) of pre-treatment is positioned at infrared image C.

3, according to claim 1 based on the infrared image background inhibition method of not having the supervision kernel regression analysis, it is as follows to it is characterized in that calculating in the step 2 step of not having supervision nuclear regression equation:

The recurrence estimation formulas is as follows:

y _i＝z(x _i)+ε _i，i＝1，……，P， (1)

Wherein, x _iBe the 2x1 dimensional vector, the coordinate of expression two-dimensional space, y _iRepresent corresponding gradation of image value; Z (x _i) be called regression function, ε _iBe stochastic error or random disturbance; With z (x _i) launch at neighborhood, can obtain following formula:

$z\left(x_i\right)={\mathrm{\β}}_0+{\mathrm{\β}}_1^T(x_i-x)+{\mathrm{\β}}_2^T\mathrm{vech}\left\{(x_i-x){(x_i-x)}^T\right\}+\·\·\·---\left(2\right)$

Definition vech () handles the vectorization of triangular portions under the symmetric matrix:

$\mathrm{vech}\left(\left[\begin{array}{cc}a& b\\ b& d\end{array}\right]\right)={\left[\begin{array}{ccc}a& b& d\end{array}\right]}^T---\left(3\right)$

β ₀=z (x), β ₁And β ₂Satisfy:

${\mathrm{\β}}_1=\▿z\left(x\right)={[\frac{\∂z\left(x\right)}{{\∂x}_1},\frac{\∂z\left(x\right)}{{\∂x}_2}]}^T---\left(4\right)$

${\mathrm{\β}}_2=\frac{1}{2}{[\frac{{\∂}^{2}z\left(x\right)}{{\∂x}_1^2},2\frac{{\∂}^{2}z\left(x\right)}{{\∂x}_1{x}_2},\frac{{\∂}^{2}z\left(x\right)}{{\∂x}_2^2}]}^T---\left(5\right)$

And parameter beta _nObtain by finding the solution following optimization problem:

$\underset{\left\{{\mathrm{\β}}_n\right\}}{\min }\underset{i=1}{\overset{P}{\mathrm{\Σ}}}{[y_i-{\mathrm{\β}}_0-{\mathrm{\β}}_1^T(x_i-x)-{\mathrm{\β}}_2^T\mathrm{vech}\{(x_i-x){(x_i-x)}^T\}-\·\·\·]}^2\·---\left(6\right)$

K _H(x _i-x，y _i-y)

Wherein

$K_H\left(t\right)=\frac{K\left(H{}^{1}t\right)}{\det \left(H\right)}---\left(7\right)$

K _H(t) be the nuclear weighting function, H is called smoothing matrix;

Utilize the mathematical operation method to carry out abbreviation, the zeroth order estimated value of trying to achieve z (x) is:

$\hat{z}\left(x\right)={\widehat{\mathrm{\β}}}_0=e_1^T{\left(X_x^T{W}_x{X}_x\right)}^{-1}{X}_x^T{W}_{x}y---\left(8\right)$

E wherein ₁ ^TBe that one first row element is 1, other is 0 column vector,

y＝[y ₁，y ₂，…，y _P] ^T (9)

W _x＝diag[K _H(x ₁-x)，K _H(x ₂-x)，…，K _H(x _P-x)] (10)

$X_x=\left[\begin{array}{ccccc}1& {(x_1-x)}^T& {\mathrm{vech}}^T\left\{(x_1-x){(x_1-x)}^T\right\}& \·\·\·& \\ 1& {(x_2-x)}^T& {\mathrm{vech}}^T\left\{(x_2-x){(x_2-x)}^T\right\}& & \·\·\·\\ \·& \·& \·& \·& \\ \·& \·& \·& \·& \\ \·& \·& \·& \·& \\ 1& {(x_P-x)}^T& {\mathrm{vech}}^T\left\{(x_P-x){(x_P-x)}^T\right\}& & \·\·\·\end{array}\right]---\left(11\right)$

By top formula as can be seen, the result of evaluation

Depend in part on the selection of smoothing matrix H; Smoothing matrix H by model representation is:

H _i＝hu _iI (12)

In the following formula, u _iCharacterization data sampling dense degree, h is called smoothing parameter.

4, according to claim 1 based on the infrared image background inhibition method of not having the supervision kernel regression analysis, it is characterized in that window (B) in current in the step 5 center test sample book (O) if background clutter, then the gray-scale value of center test sample book (O) will be removed.

5, according to claim 1 based on the infrared image background inhibition method of not having the supervision kernel regression analysis, it is characterized in that the center test sample book (O) of window (B) in current in the step 5 is if object pixel then suppresses to be superimposed upon the gray-scale value of the background clutter on the center test sample book (O) of interior window (B).

Images