CN102314687A - Method for detecting small targets in infrared sequence images - Google Patents

Abstract

The invention discloses a method for detecting small targets in infrared sequence images, which belongs to the field of image data processing. The method includes the following steps: (1) image preprocessing step: obtaining the differential image of each frame of image of an obtained infrared sequence image; (2) background modeling step: obtaining the characteristic image of each frame of image; (3) smallest non-uniform image segmentation step: working out an optimal segmentation threshold value, and utilizing the optimal segmentation threshold value to segment the obtained characteristic image, so that a small infrared target can be detected. For an image with a uniformly distributed background, the method can effectively rebuild the background; the smallest non-uniform image segmentation method converts the segmentation problem into the segmentation problem of images, the optimal segmentation threshold value is sought from the perspective of segmentation energy, and targets can be more accurately segmented out.

Description

Small target detecting method in a kind of infrared sequence image

Technical field

The invention belongs to image processing method, be specifically related to the detection method of infrared small target under a kind of complex background.

Background technology

Infrared small target can be described as a relative concept, and so-called " little " is meant that its physical dimension is little in image, it has been generally acknowledged that little target is meant in the target of size between 1*1 pixel to 10*10 pixel on the image.Because target is less, there is not the shape information that can be used for discerning, it is lower also to have signal to noise ratio (S/N ratio) in addition,, the characteristics that position and movement rate are all unknown make the detection of infrared small target become very difficult.The infrared small target detection algorithm that exists at present mainly comprises the detection algorithm based on wave filter, based on the detection algorithm of wavelet transformation, based on the detection algorithm of mathematical morphology and some are based on detection algorithm study or movable information.Detection algorithm based on wave filter leaches possible target area through wave filter, suppresses most of background, and these class methods are simply effective, but very little in target, and effect is bad under the lower situation of signal to noise ratio (S/N ratio), effectively filtering interfering; Detection algorithm based on wavelet transformation is that different characteristic detects between target area and the background through extracting, but under the less situation of target, small echo can not effectively extract the characteristic of target usually; Method based on mathematical morphology mainly is to use the top-hat conversion to suppress background, but it is very sensitive to choosing of structural element, and is in addition, also very sensitive to the extraction of little target under the low signal-to-noise ratio; Method based on study is only obtaining under the situation of training sample preferably good detection effect to be arranged, but usually preferably obtaining of training sample be difficulty very.In a word, all there is certain limitation in a series of detection algorithms that exist at present.

Summary of the invention:

The present invention proposes the small target detecting method in a kind of infrared sequence image, at first rebuild and remove background, adopt the method for minimum lack of balance figure cutting then, utilize image segmentation that little target is detected, can be partitioned into target accurately through background.。

The method of the little target detection of a kind of infrared sequence image of the present invention, step is:

(1) image pre-treatment step:

For better detection effect is arranged, at first to carry out pre-service, to reach the purpose that single frames strengthens target to image; At first each row of original image is removed this row average; Respectively image is carried out mean filter with two windows that differ in size then, calculate its difference then, obtain error image.

(2) background modeling step:

From neighborhood of pixels, at first to any adjacent two frame error image f ₁And f ₂Do piecemeal respectively and handle, with f ₁In certain block of pixels be regarded as a vector, from f ₂In k block of pixels of its arest neighbors of search, arest neighbors estimate the use Euclidean distance, remove first block of pixels nearest with it, then, under the condition that satisfies minimum reconstruction error, with k-1 remaining this block of pixels of block of pixels reconstruction, to f ₁In each block of pixels do identical processing, obtain reconstructed image at last; By error image f ₁Subtract each other with reconstructed image, obtain characteristic image at last;

The basis of doing like this is; Background is equally distributed comparatively speaking, and each background piece has many block of pixels in image, and target is then less; Background can well be rebuild; The error that target is then rebuild is bigger, and during as characteristic image, then target can highlight with the reconstruction error image.

(3) minimum lack of balance figure cutting step:

At first set up grid chart model G between pixel (V, E), wherein node is a pixel, the limit be between neighbor constraint w (u, v), (u v) is used to define the flatness of neighbor to this constraint w;

Grid model is regarded as three-dimensional plot, and then segmentation problem is converted into the cutting problem of figure, highly is pixel grey scale; Planimetric coordinates is a locations of pixels, and cut surface can be regarded as the threshold value t of image segmentation, and threshold value t is divided into A with image; B two parts calculate A, ablation energy between B and optimization function size; Down carry out searching loop from the maximal value of t, make optimization function get the t of minimum value, be optimal segmenting threshold; With this segmentation threshold the above-mentioned characteristic image that obtains is cut apart then, thereby detected infrared small target.

Wherein, said background modeling step process is:

(2.1) for any two adjacent two field picture f ₁And f ₂, respectively it is done piecemeal and handles;

(2.2) to f ₁In each block of pixels, it is expressed with vector form, at frame f ₂K block of pixels of middle search arest neighbors;

(2.3) for f ₁In each block of pixels, except f ₂In first block of pixels nearest with it, by it at f ₂Middle corresponding other k-1 block of pixels is rebuild according to reconstruction formula, and satisfies minimum reconstruction error, and wherein said reconstruction formula is:

${\left|\right|B_1^i-\mathrm{\Σ}{w}_j{B}_2^j\left|\right|}_2,j=1,...,k-1$

Wherein, Expression f ₁In i block of pixels,

Expression f ₂In j nearest neighbor pixels piece, w _jIt is the weight coefficient of j nearest neighbor pixels piece;

(2.4) f ₁In the reconstructed block value of i block of pixels be taken as

The average of the reconstructed value of two pieces, the f that asks are got in overlapping region between two pieces ₁In the reconstructed block value of each block of pixels after, the background image f that can obtain rebuilding _r

(2.5) last, image f ₁The background image f that obtains with reconstruction _rDo difference, promptly obtain characteristic image f ₁-f _r, then removed most of background in the characteristic image of this moment, target can highlight.

Wherein, said minimum lack of balance figure cutting step process is:

(3.1) at first set up grid chart model G between pixel (V, E), wherein node is a pixel, the limit be between neighbor constraint w (u, v), (u v) is used to define the flatness of neighbor to this constraint w; The grid chart model is regarded as three-dimensional plot, and then segmentation problem is converted into the cutting problem of figure, wherein highly is pixel grey scale, and planimetric coordinates is a locations of pixels, and cut surface can be regarded as the threshold value t of image segmentation; The node that image segmentation is about among the figure is divided into A, and B two parts are higher than the A that is on the cut surface, are lower than the B that is of cut surface.

(3.2) ablation energy between calculating A and B defines as follows:

$\mathrm{Cut}(A,B)=\underset{u\∈A,v\∈B}{\mathrm{\Σ}}w(u,v)$

Desire accurately to cut apart little object boundary, reach the purpose of minimal cut, then require A and B to cut apart in the minimum position of constraint, (A B) measures then to use ablation energy Cut; Because mainly be to choose segmentation threshold here, so energy calculating is only relevant with threshold value t, promptly above-mentioned energy computing formula can be rewritten as:

$\mathrm{Cut}(A,B)=\underset{I\left(u\right)\&GreaterEqual;t,I\left(v\right)<t}{\mathrm{\&Sigma;}}w(u,v)$

Wherein, I (u) is the gray-scale value of pixel u, I (v) be the gray-scale value of pixel v, t is current threshold value, wherein w (u v) is the constraint between current pixel point u and its neighborhood territory pixel point v, and formula is following:

$w(u,v)=\left\{\begin{array}{cc}\exp -\left[\right||I\left(u\right)-I\left(v\right)||+||u-v|\left|\right],& \left|\right|u-v\left|\right|<D\\ 0,& \mathrm{otherwise}\end{array}\right.$

‖ u-v ‖＜D remarked pixel point u in the formula and the distance between the pixel v are less than D; Generally v is the pixel of eight neighborhoods of u, can certainly get bigger scope, and the size of D depends on the size of selected neighborhood; Neighborhood varies in size, and distance B is also different.

(3.3) cut principle based on minimum lack of balance figure, can be constructed as follows optimization function:

$L(A,B)=\frac{\mathrm{Cut}(A,B)}{\left|A\right|}+\frac{\mathrm{Cut}(A,B)}{\left|B\right|}$

Wherein | A|, | B| is respectively the sum of all pixels of A, B;

From maximum t, down search always, (A B) reaches minimum value and promptly stops L, and this moment, the value of t was optimal segmenting threshold.

The method of the background modeling that the present invention proposes is the angle from neighborhood of pixels, for certain block of pixels, considers to rebuild this block of pixels from the plurality of pixels piece of another frame; And satisfy minimum reconstruction error, and this method is for the relative uniform image of background distributions, and each background piece has many block of pixels in image; Target is then less; So background can well be rebuild, and the error that target is rebuild is bigger, thereby well reaches the purpose of removing background; Minimum lack of balance figure cutting method is converted into the cutting problem of figure with segmentation problem, from the angle of ablation energy, goes to seek best segmentation threshold, can be partitioned into target more accurately.

Description of drawings

Fig. 1 is a process flow diagram of the present invention;

Fig. 2 (a) is the former figure that comprises the ground infrared small target;

Fig. 2 (b) is pretreated result;

Fig. 2 (c) is result after the removal background;

Fig. 2 (d) is the result of minimum lack of balance figure cutting;

Fig. 3 (a) is the former figure that comprises sea day infrared small target;

Fig. 3 (b) is pretreated result;

Fig. 3 (c) is result after the removal background;

Fig. 3 (d) is the result of minimum lack of balance figure cutting.

Embodiment

Below in conjunction with accompanying drawing and specific embodiment the present invention is done further explain.

The method of the little target detection of a kind of infrared sequence image of the present invention, concrete steps are:

(1) image pre-service

In order to reach better detection effect, need earlier to image graph 2 (a) pre-service, to reach the purpose that single frames strengthens target.At first each row of original image is removed this row average; Adopt the window of 15*15 and the window of 3*3 respectively the image after the removal average to be carried out mean filter then; Calculate filtered image difference at last twice, the error image Fig. 2 (b) that obtains is pretreated image.

2, background modeling is removed background

(2.1) to the pretreated image f of adjacent two frames ₁And f ₂Carry out piecemeal, the size of the piece of choosing is 11*11 here, and the overlapping size of interblock is 2.

(2.2) to f ₁In each block of pixels, it is expressed with vector form, at frame f ₂K block of pixels of middle search arest neighbors, common Euclidean distance is used in estimating of arest neighbors.

(2.3) for fear of the piece that comprises target is directly rebuild by the object block of correspondence, for f ₁In each block of pixels, except f ₂In first block of pixels nearest with it, by it at f ₂Middle corresponding other k-1 block of pixels is rebuild according to reconstruction formula, and satisfies minimum reconstruction error, and wherein said reconstruction formula is:

${\left|\right|B_1^i-\mathrm{\&Sigma;}{w}_j{B}_2^j\left|\right|}_2,j=1,...,k-1$

Wherein,

Expression f ₁In i block of pixels,

Expression f ₂In j nearest neighbor pixels piece, w _jBe the weight coefficient of j nearest neighbor pixels piece, this formulate calculates two norms, and satisfies the condition of least error.

(2.4) f ₁In the reconstructed block value of i block of pixels be taken as The average of the reconstructed value of two pieces, the f that asks are got in overlapping region between two pieces ₁In the reconstructed block value of each block of pixels after, the background image f that can obtain rebuilding _r

(2.5) last, image f ₁The background image f that obtains with reconstruction _rDo difference, promptly obtain characteristic image f ₁-f _r, then removed most of background in the characteristic image of this moment, target can highlight, shown in Fig. 2 (c).

3, minimum lack of balance figure cutting

(3.1) at first set up grid chart model G between pixel (V, E), wherein node is a pixel, the limit be between neighbor constraint w (u, v), (u v) is used to define the flatness of neighbor to this constraint w; The grid chart model is regarded as three-dimensional plot, and then segmentation problem is converted into the cutting problem of figure, wherein highly is pixel grey scale, and planimetric coordinates is a locations of pixels, and cut surface can be regarded as the threshold value t of image segmentation; The node that image segmentation is about among the figure is divided into A, and B two parts are higher than the A that is on the cut surface, are lower than the B that is of cut surface.

(3.2) to current threshold value t, according to formula calculate current ablation energy Cut (A, B), pixel u wherein, the pixel value of v satisfy I (u) >=t and I (v)＜t, u, v get eight neighborhoods,

$\mathrm{Cut}(A,B)=\underset{I\left(u\right)\&GreaterEqual;t,I\left(v\right)<t}{\mathrm{\&Sigma;}}w(u,v)$

W (u v) is the constraint between current pixel point u and its neighborhood territory pixel point v, and formula is following:

$w(u,v)=\left\{\begin{array}{cc}\exp -\left[\right||I\left(u\right)-I\left(v\right)||+||u-v|\left|\right],& \left|\right|u-v\left|\right|<D\\ 0,& \mathrm{otherwise}\end{array}\right.$

‖ u-v ‖＜D remarked pixel point u in the formula and the distance between the pixel v are less than D, and v is the pixel of eight neighborhoods of u here, and the size of D depends on the size of selected neighborhood, and neighborhood varies in size, and distance B is also different.

(3.3) statistics is greater than the pixel summation of the A of threshold value t part, as | the value of A|; Statistics is less than the number summation of the pixel of the B of threshold value t part, as | the value of B|, same, u, v get 8 neighborhoods.

(3.4) the Cut that obtains (A, B), | A|, | the value substitution optimization function of B|, calculate L (the optimization function formula is for A, B) size:

$L(A,B)=\frac{\mathrm{Cut}(A,B)}{\left|A\right|}+\frac{\mathrm{Cut}(A,B)}{\left|B\right|}$

(3.5) above step is carried out in circulation, and threshold value t down searches for from maximal value always; Make L (A; B) get the threshold value t of minimum value, be best segmentation threshold, resulting characteristic image in the 2nd step is cut apart with this threshold value; Can be partitioned into little target, last segmentation result such as Fig. 2 (d).

The infrared small target that method of the present invention can be applied under the multiple image background detects, and like infrared small target detection under day background of sea etc., its treatment step is same as the previously described embodiments.Fig. 3 (a) is the former figure that comprises sea day infrared small target; The pretreated result of Fig. 3 (b); Result after Fig. 3 (c) removal background; The result of the minimum lack of balance figure cutting of Fig. 3 (d).

Claims (5)

1. the method for the little target detection of infrared sequence image comprises following process:

(1) image pre-treatment step:

For each two field picture in the infrared sequence image; At first, remove the average of each row in the image, respectively image is carried out mean filter with two windows that differ in size then; The last difference of computed image behind twice mean filter obtains the error image of each two field picture;

(2) background modeling step is promptly set up the reconstructed image of the error image of each two field picture, and obtains the characteristic image of each two field picture, is specially:

At first, the error image of the consecutive frame image of the error image of each two field picture and this each two field picture being done piecemeal respectively handles;

Secondly, the arbitrary block of pixels with the error image of this each two field picture is regarded as vector, k block of pixels of its arest neighbors of search from another error image; Remove a block of pixels nearest with it; Then according to k-1 remaining this block of pixels of neighbour's block of pixels reconstruction,, can obtain the reconstructed image of this error image to the reconstruction of all block of pixels of error image of this each frame; Wherein, k is the positive integer greater than 1;

Then, error image and its reconstructed image of this each two field picture subtracted each other, obtain the characteristic image of each two field picture;

(3) minimum lack of balance figure cutting step:

For the characteristic image of each two field picture, at first set up grid chart model G between pixel (V, E), wherein node is a pixel, the limit be between neighbor constraint w (u, v), (u v) is used to define the flatness of neighbor to this constraint w;

Secondly, (V E) is regarded as three-dimensional plot with grid model G; Wherein highly be pixel grey scale, planimetric coordinates is a locations of pixels, with threshold value t as cut surface; Node in the three-dimensional plot is divided into A and B two parts, and ablation energy between computed image A and image B two parts and optimization function are big or small, and down carry out searching loop from the maximal value of threshold value t; Make optimization function get the threshold value of minimum value, as optimal segmenting threshold;

At last, utilize this optimal segmenting threshold that the characteristic image of above-mentioned each two field picture that obtains is cut apart, can be partitioned into infrared small target, accomplish and detect.

2. method according to claim 1 is characterized in that, in the background modeling of said step (2), the reconstruction formula of said this block of pixels of reconstruction is:

${\left|\right|B_1^i-\mathrm{\&Sigma;}{w}_j{B}_2^j\left|\right|}_2,j=1,...,k-1$

Wherein,

For waiting to rebuild the error image f at block of pixels place ₁In i block of pixels,

The error image f of expression consecutive frame ₂In j nearest neighbor pixels piece, w _jBe the weight coefficient of this j block of pixels, i, j are positive integer, ‖ ₂Two norms are calculated in expression, and satisfy the condition of least error.

3. method according to claim 2 is characterized in that, estimating of said arest neighbors is Euclidean distance.

4. according to the described method of one of claim 1-3, it is characterized in that in the described step (3), the ablation energy between image A and image B defines as follows:

$\mathrm{Cut}(A,B)=\underset{u\&Element;A,v\&Element;B}{\mathrm{\&Sigma;}}w(u,v)$

Above-mentioned energy computing formula can be rewritten as:

$\mathrm{Cut}(A,B)=\underset{I\left(u\right)\&GreaterEqual;t,I\left(v\right)<t}{\mathrm{\&Sigma;}}w(u,v)$

Wherein, I (u) is the gray-scale value of pixel u, I (v) be the gray-scale value of pixel v, t is current threshold value, w (u v) is the constraint between current pixel point u and its neighborhood territory pixel point v, and its computing formula is following:

$w(u,v)=\left\{\begin{array}{cc}\exp -\left[\right||I\left(u\right)-I\left(v\right)||+||u-v|\left|\right],& \left|\right|u-v\left|\right|<D\\ 0,& \mathrm{otherwise}\end{array}\right.$

‖ u-v ‖＜D remarked pixel point u in the formula and the distance between the pixel v are less than D, and D is a positive integer.

5. according to the described method of one of claim 1-4, it is characterized in that in the described step (3), said optimization function is:

$L(A,B)=\frac{\mathrm{Cut}(A,B)}{\left|A\right|}+\frac{\mathrm{Cut}(A,B)}{\left|B\right|}$

Wherein | A|, | B| is respectively the sum of all pixels of image A and image B.

Images