CN104268835A - Weak and small infrared target strengthening method based on image geometric separation - Google Patents

Abstract

The invention relates to a weak and small infrared target strengthening method based on image geometric separation. Firstly, according to the distribution characteristics of weak and small targets and background clutters in an infrared image, a spectrogram wavelet transform dictionary and a non-subsampled shearlet transform dictionary are adopted for carrying out sparse representation on the weak and small targets and the background clutter constituents in the image respectively; then, the two unrelated sparse dictionaries are introduced to a image geometric separation framework, total variation penalty factors are utilized for obtaining more background clutter constituent information so that images can be separated more effectively, and therefore the purpose of strengthening weak and small target signals is achieved. The method has the better effects of strengthening the weak and small targets and suppressing the background clutters; compared with classical filtering methods such as the maximum mean method, the morphological filtering method and the two-dimensional minimum mean square error method, background suppression factors are improved, and false alarms caused by strong edges are well reduced.

Description

A kind of infrared small object Enhancement Method be separated based on image geometry

Technical field

The invention belongs to infrared image processing field, be specifically related to the Enhancement Method of infrared small object in a kind of complex background.

Background technology

In infrared technique application, Dim targets detection is a very crucial technology in infrared track recognition system, infraed early warning system and precise guidance system.Due to the complicated abominable of infrared imaging environment and the limitation of detector hardware condition, infrared image contrast and the signal to noise ratio (S/N ratio) of acquisition are all very low, in addition when target range infrared detection system is relatively far away, target on the detector imaging area is less, substantially present point-like, the background clutter that therefore target is easy to by strongly rising and falling floods.

At present, for this technical barrier of small IR targets detection, domestic and international scientific research scholar has carried out many research work, and has made some progress.Conventional classical way has Largest Mean/medium filtering, and its structural design is easy, and implementation procedure is relatively simple, and background suppress is respond well, and where the shoe pinches is in the selection of filter window size; Morphology Top-Hat filtering is another conventional Image Enhancement, has stronger practicality, can effective Background suppression, but finds out the difficult point place that suitable structural element is the method; Two dimension least mean-square error filtering (TDLMS) method is a kind of typical adaptive line background prediction methods, and its structure is simple, speed, but can be deteriorated for non-stationary background (as cloud layer) inhibition.

For the above-mentioned detection method based on background forecast to the deficiency in Dim targets detection ability, sparse signal representation theory is used in the detection to infrared small object by some document.At published patent of invention " a kind of method for detecting infrared puniness target based on shearing wave conversion " (grant number: CN 102324021 B, inventor: Peng Zhenming, Wei Fang, Peng Lingbing etc.) middle employing shearing wave rarefaction representation infrared image, then Weak target is utilized to be in this feature of high-frequency information figure, high-frequency information figure is carried out to the process of background suppress and Weak target enhancing, Threshold segmentation is finally adopted to extract Weak target, obtain good Detection results, but some are existed to the complex background of strong edge contour feature, more background clutter composition can be there is in high-frequency information figure after process, inhibition shows slightly not enough, and at document " small IR targets detection based on anatomic element rarefaction representation " (see " playing arrow and guidance journal ", 33 volumes (4 phase): P29-32 in 2013, author: Li Zhengzhou, Wang Huigai, Liu Mei, Ding Hao, Jin Gang) in propose a kind of detection method of small target of the self-adapting signal rarefaction representation based on image aspects constituent analysis theory.First the method carries out K cluster svd (K_SVD) to infrared image, train adaptive rarefaction representation dictionary, then threshold function table is adopted to distinguish the two seed dictionaries representing target and background composition, finally threshold decision is carried out to the sparse coefficient of the sub-dictionary of target and determine target region, complete the detection to Weak target.The method utilizes threshold function table to distinguish to same sparse dictionary in the selection of two seed dictionaries representing target and background composition, therefore the correlativity of the two is very large, with the infrared image background set in eight-legged essay be change slowly, not there is the uniform background environment (as deep space) of complex edge and grain details, these factors cause the adaptability of the method to complex background environment poor, limited to the detectability of Weak target.

Summary of the invention

The object of the invention is to the deficiency for existing method for detecting infrared puniness target, a kind of infrared small object Enhancement Method be separated based on image geometry is provided.Image geometry separation theorem in sparse signal representation is applied to infrared small object and strengthens by the method, thus realizes the enhancing at infrared small target in complex background, effective Background suppression clutter.

Solution of the present invention is first according to Weak target in infrared image and background clutter characteristic distributions, adopts the shearing wave of spectrogram wavelet transformation and non-lower sampling to convert dictionary respectively carry out rarefaction representation to Weak target in image and background clutter composition; Then these two mutual incoherent sparse dictionary are incorporated in image geometry split-frame, and utilize full variation penalty factor to obtain more background clutter composition informations, make it obtain more effective separation, thus reach the object strengthening Weak target signal.

The step of the inventive method specifically describes as follows:

(1) infrared image separated modeling initialization;

A. infrared image separated modeling expression formula is

$\begin{array}{c}\{{\mathrm{\α}}_T^{\mathrm{opt}},{\mathrm{\α}}_B^{\mathrm{opt}}\}=\underset{\{{\mathrm{\α}}_T,{\mathrm{\α}}_B\}}{\arg \min }{\left|\right|{\mathrm{\α}}_T\left|\right|}_1+{\left|\right|{\mathrm{\α}}_B\left|\right|}_1+\mathrm{TV}\left\{D_B{\mathrm{\α}}_B\right\}\end{array}s.t.X=D_T{\mathrm{\α}}_T+D_B{\mathrm{\α}}_B;$

B. initialization X _b=X and X _t=X-X _b;

(2) Geometrically split is carried out to infrared image;

A. X is fixed _t, upgrade X _b;

---to X _bcarry out NSST conversion:

---to factor alpha _bobtain with soft threshold method

---reconstruct

B. penalty factor TV is increased;

---carry out TV correction: ${\hat{X}}_B^{\′}={\hat{X}}_B-\mathrm{\μ}\frac{\∂\mathrm{TV}\left\{{\hat{X}}_B\right\}}{\∂{\hat{X}}_B}={\hat{X}}_B-\mathrm{\μ}\▿\·\left(\frac{\▿{\hat{X}}_B}{|\▿{\hat{X}}_B|}\right);$

C. X is fixed _b, upgrade X _t;

---to X _tcarry out SGWT conversion:

---to factor alpha _tobtain with soft threshold method

---reconstruct

(3) the rear Weak target image of enhancing is exported.

The present invention adopts the shearing wave of spectrogram wavelet transformation and non-lower sampling conversion dictionary to carry out rarefaction representation to Weak target in image and background clutter respectively.Because spectrogram wavelet transformation has stronger sub-band division ability compared with Morlet wavelet transform, and it is good to the carving effect of local singular value, represent that Weak target composition not only can make Weak target and background clutter composition distinguish and come with it, can effectively keep Weak target information simultaneously; And background clutter composition advantage is that the conversion of non-lower sampling shearing wave has good time-frequency locality to adopt the conversion of non-lower sampling shearing wave to represent, the directional sensitivity of height, and the ability of close optimum expression two dimensional image signal, especially to portraying of the wire singular value existed in image, there is good expression effect.By two groups of emulation experiments, effectively demonstrate these two kinds of sparse transformation to Weak target and the close optimum discrete representation effect of background clutter composition, and find out from the result figure (Fig. 4) of experiment 2, the separation inhibition of this method to the background component with strong edge features such as sea horizons is especially good, effectively highlights Weak target.

Accompanying drawing explanation

Fig. 1 chooses scene and dimensional strength figure thereof in experiment 1;

Fig. 2 is enhancing result and the three-dimensional plot thereof of several method in experiment 1.(a) and (e) Largest Mean method result and three-dimensional plot thereof; (b) and (f) Top-hat method result and three-dimensional plot thereof; (c) and (g) TDLMS method result and three-dimensional plot thereof; (d) and (h) the inventive method result and three-dimensional plot thereof;

Fig. 3 chooses scene and dimensional strength figure thereof in experiment 2;

Fig. 4 is enhancing result and the three-dimensional plot thereof of several method in experiment 2.(a) and (e) Largest Mean method result and three-dimensional plot thereof; (b) and (f) Top-hat method result and three-dimensional plot thereof; (c) and (g) TDLMS method result and three-dimensional plot thereof; (d) and (h) the inventive method result and three-dimensional plot thereof.

Embodiment

Below the implementation step of the inventive method is described in further detail:

(1) separated modeling is carried out and initialization to the infrared image comprising Weak target;

(1a) infrared image separated modeling:

First image geometry separating thought is expressed as follows: suppose that composition diagram is come from the individual different independent source of n as the informational content of I, so just think that this image is made up of n independent sector, i.e. I=I ₁+ I ₂+ I ₃+ ... + I _n, often kind of informational content I simultaneously _i(i ∈ 1 ..., n) all by a corresponding with it conversion dictionary Di rarefaction representation, can be expressed as

${\mathrm{\α}}_i^{\mathrm{opt}}=\arg \min {\left|\right|{\mathrm{\α}}_i\left|\right|}_{0}s.t.I_i=D_i{\mathrm{\α}}_i---\left(1\right)$

By adopting a said n mutual incoherent dictionary D, can by each different constituent I of image I _i(i ∈ 1 ..., n) near-optimization ground rarefaction representation out, thus reach the object of image geometry separation.

According to image geometry separating thought, can the infrared image X comprising Weak target be regarded as Weak target X _twith background clutter X _blinear superposition, be expressed as:

X＝X _T+X _B (2)

Then the Geometrically split of infrared image can be expressed as:

$\begin{array}{c}\{{\mathrm{\α}}_T^{\mathrm{opt}},{\mathrm{\α}}_B^{\mathrm{opt}}\}=\underset{\{{\mathrm{\α}}_T,{\mathrm{\α}}_B\}}{\arg \min }{\left|\right|{\mathrm{\α}}_T\left|\right|}_0+{\left|\right|{\mathrm{\α}}_B\left|\right|}_0\end{array}s.t.X=D_T{\mathrm{\α}}_T+D_B{\mathrm{\α}}_B---\left(3\right)$

Wherein, D _tand D _brepresent the conversion dictionary of energy Its Sparse Decomposition target component and background clutter composition respectively, D _tα _t≈ X _t, D _bα _b≈ X _brepresent the sparse bayesian learning of target component and background clutter composition respectively; α _tand α _brepresent that target component and background clutter composition are at dictionary D respectively _tand D _bin sparse bayesian learning coefficient, utilize two kinds to convert dictionaries and carry out Its Sparse Decomposition and obtain image sparse coefficient with

In view of solving l ₀the complicacy of norm and difficulty, usually with solving l ₁norm replaces solving l ₀, loosely become a convex optimization linear problem by optimization problem.Simultaneously because full variation TV method can receive very good effect in the expression containing limbus characteristic image, therefore increase a full variation penalty factor, make background clutter composition obtain more effective separation.Expression formula is written as:

$\begin{array}{c}\{{\mathrm{\α}}_T^{\mathrm{opt}},{\mathrm{\α}}_B^{\mathrm{opt}}\}=\underset{\{{\mathrm{\α}}_T,{\mathrm{\α}}_B\}}{\arg \min }{\left|\right|{\mathrm{\α}}_T\left|\right|}_1+{\left|\right|{\mathrm{\α}}_B\left|\right|}_1+\mathrm{TV}\left\{D_B{\mathrm{\α}}_B\right\}\end{array}s.t.X=D_T{\mathrm{\α}}_T+D_B{\mathrm{\α}}_B---\left(4\right)$

The full variation TV (I) of image I represents the l of image gradient in essence ₁norm.Background image D can be made as penalty factor with TV _bα _bthere is sparse gradient, thus retain more background clutter composition informations.

(1b) initialization;

Weak target component-part diagram picture and background clutter component-part diagram picture can be expressed as X _b=X and X _t=X-X _b.

(2) Geometrically split is carried out to infrared image;

(2a) fixing X _t, upgrade X _b;

First choose non-lower sampling shearing wave conversion (NSST) and carry out rarefaction representation background clutter composition, its expression formula is:

${\mathrm{\α}}_B=D_B^{*}{X}_B---\left(5\right)$

Wherein, α _brepresent the sparse coefficient after conversion;

Then to factor alpha _bsoft threshold method process is adopted to obtain the coefficient of approximate representation background clutter composition

Finally reconstruct background clutter component-part diagram picture be expressed as:

${\hat{X}}_B=D_B{\widehat{\mathrm{\α}}}_B---\left(6\right)$

Background clutter composition advantage is that the conversion of non-lower sampling shearing wave has good time-frequency locality to adopt NSST to represent, directional sensitivity to heavens, and the ability of close optimum expression two dimensional image signal, especially to portraying of the wire singular value existed in image, there is good expression effect.

(2b) penalty factor TV is increased;

Adopt the full variation factor to correct to isolated background clutter composition, expression formula is:

${\hat{X}}_B^{\′}={\hat{X}}_B-\mathrm{\μ}\frac{\∂\mathrm{TV}\left\{{\hat{X}}_B\right\}}{\∂{\hat{X}}_B}={\hat{X}}_B-\mathrm{\μ}\▿\·\left(\frac{\▿{\hat{X}}_B}{|\▿{\hat{X}}_B|}\right)---\left(7\right)$

Wherein, μ is regulating parameter.

The full variation TV of image represents the l of image gradient in essence ₁norm.Background clutter image can be made with TV to have sparse gradient as penalty factor, thus effectively get rid of Weak target composition, optimize background clutter composition information.

(2c) fixing X _b, upgrade X _t;

First adopt spectrogram wavelet transformation to carry out rarefaction representation Weak target composition, it is expressed as:

${\mathrm{\α}}_T=D_T^{*}{X}_T---\left(8\right)$

Wherein, α _trepresent the sparse coefficient after conversion;

Then to factor alpha _tsoft threshold method process is adopted to obtain the coefficient of approximate representation Weak target composition

Finally reconstruct Weak target component-part diagram picture be expressed as:

${\hat{X}}_T=D_T{\widehat{\mathrm{\α}}}_T---\left(9\right)$

Spectrogram wavelet transformation (Spectral Graph Wavelet Transform, SGWT) is incorporated into by spectral graph theory in wavelet transformation thus realizes the emerging signal decomposition instrument of one of sparse signal representation.It inherits many advantageous properties of Morlet wavelet transform, as frequency localization characteristic, multiscale analysis characteristic etc., possesses the ability of the angle analysis signal characteristic from figure simultaneously, compared with wavelet transformation, it has better performance in decomposed signal to different sub-band, is widely used in image procossing.Given this, the present invention adopts spectrogram wavelet transformation to carry out rarefaction representation Weak target composition, effectively portrays point-like singular value, fully keeps Weak target information.

(3) the rear Weak target image of enhancing is exported.

Because background clutter composition obtains effective separation and suppression, the Weak target image enhancement effects of acquisition is remarkable, can contrast and find out from the design sketch of two groups of emulation experiments (Fig. 2 and Fig. 4).

Two groups of emulation experiments are carried out to the inventive method, the Image Enhancement of several classics introduced before choosing, the filtering method methods as a comparison such as Largest Mean, morphology Top-Hat operator and TDLMS.

Experiment 1: as shown in Fig. 1 (a), choose the true Infrared DIM-small Target Image that a width take cloud layer as background, its size is 128 × 128, and (b) is the three-dimensional plot of this image, therefrom can find out that Weak target is submerged in high strength background clutter, cannot accurately identify.

As shown in Figure 2, give the enhancing result of the inventive method and the enhancing result of control methods and they each self-corresponding dimensional strength figure.Therefrom can visually see, strengthening the property of the inventive method is obviously better than other detection methods, wherein schemes (a) Largest Mean filtering method enhancing effect poor, produces false profile, cause target unintelligible, situs ambiguus around target; It is slightly better than Largest Mean filter effect that figure (b) Top-hat filtering method strengthens result, but background clutter composition is still a lot.Figure (c) TDLMS filtering method strengthens the first two method that result is better than again, but background component is still obviously visible, and produces diplopia around target, easily causes the fuzzy of target; By contrast, it is more excellent that the method that figure (d) the present invention proposes strengthens effect, successfully separated by target component, substantially do not observe background component simultaneously, embody good strengthening the property.

Experiment 2: as shown in Fig. 3 (a), choose the sea Weak target infrared image that a width comprises sea horizon background component, its size is 128 × 128, and (b) is the dimensional strength figure of this image.Obviously visible, Weak target is near sea horizon, and sea clutter background rises and falls violent simultaneously, for the detection identification of Weak target brings severe jamming and difficulty.

As shown in Figure 4, give the enhancing result of the inventive method and the enhancing result of control methods and they each self-corresponding dimensional strength figure.As can be seen from the figure, for the infrared image containing limbus features such as sea horizons, the detection perform of the inventive method is still better than other Enhancement Method.Wherein scheme (a) Largest Mean filtering method and cause object and background aliasing, there is a large amount of wire background component simultaneously, strengthen effect poor; Figure (b) Top-hat filtering method is pretty good to wire background suppress effect, but sea clutter background composition is still obviously visible and gray scale is higher; The obvious weak point of figure (c) TDLMS filtering method easily produces artifact to the edge contour of Weak target, sea horizon etc.; By contrast, the method Detection results that figure (d) the present invention proposes is better, is effectively separated by the clutter background composition comprising sea horizon, thus obtains the image only containing target component, embodies excellent enhancing effect.

In order to further illustrate the relative good performance of institute of the present invention extracting method, employing Background suppression factor (BSF) this objective parameter carrys out the quality that quantitative comparison distinct methods is strengthened the property to Weak target.The value of BSF is larger, illustrates that corresponding method keeps the ability of target component information and Background suppression clutter composition stronger.As seen from Table 1, through the inventive method, the BSF numerical value calculated after above-mentioned two groups of experiment scene process is obviously greater than to the BSF numerical value of other Enhancement Method, thus reflect that the inventive method is to the more excellent rejection ability of background clutter composition objectively, effectively realize the enhancing to infrared small object.

Table 1 distinct methods to the BSF numeric ratio after two groups of experiment scene process comparatively

The foregoing is only embodiments of the invention, be not limited to the present invention.The present invention can have various suitable change and change.All any amendments done within the spirit and principles in the present invention, equivalent replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (2)

1., based on the infrared small object Enhancement Method that image geometry is separated, it is characterized in that:

First, according to Weak target in infrared image and background clutter characteristic distributions, adopt the shearing wave of spectrogram wavelet transformation and non-lower sampling to convert dictionary respectively to Weak target in image and background clutter composition and carry out rarefaction representation; Then these two mutual incoherent sparse dictionary are incorporated in image geometry split-frame, and utilize full variation penalty factor to obtain more background clutter composition informations, make it obtain more effective separation, thus reach the object strengthening Weak target signal.

2. as claimed in claim 1 a kind of based on image geometry be separated infrared small object Enhancement Method, it is characterized in that:

Choose spectrogram wavelet transformation (SGWT) and convert (NSST) respectively as characterizing target component X with non-lower sampling shearing wave _tdictionary D _twith sign background clutter component X _bdictionary D _bcarry out Geometrically split to the infrared image X comprising Weak target, specific implementation step is as follows:

(1) infrared image separated modeling initialization:

A. infrared image separated modeling expression formula is

$\begin{array}{c}\{{\mathrm{\α}}_T^{\mathrm{opt}},{\mathrm{\α}}_B^{\mathrm{opt}}\}=\underset{\{{\mathrm{\α}}_T,{\mathrm{\α}}_B\}}{\arg \min }{\left|\right|{\mathrm{\α}}_T\left|\right|}_1+{\left|\right|{\mathrm{\α}}_B\left|\right|}_1+\mathrm{TV}\left\{D_B{\mathrm{\α}}_B\right\}\end{array}s.t.X=D_T{\mathrm{\α}}_T+D_B{\mathrm{\α}}_B;$

B. initialization X _b=X and X _t=X-X _b;

(2) Geometrically split is carried out to infrared image:

A. X is fixed _t, upgrade X _b;

---to X _bcarry out NSST conversion:

---to factor alpha _bobtain with soft threshold method

---reconstruct

B. penalty factor TV is increased;

---carry out TV correction: ${\hat{X}}_B^{\′}={\hat{X}}_B-\mathrm{\μ}\frac{\∂\mathrm{TV}\left\{{\hat{X}}_B\right\}}{\∂{\hat{X}}_B}={\hat{X}}_B-\mathrm{\μ}\▿\·\left(\frac{\▿{\hat{X}}_B}{|\▿{\hat{X}}_B|}\right);$

C. X is fixed _b, upgrade X _t;

---to X ^tcarry out SGWT conversion:

---to factor alpha _tobtain with soft threshold method

---reconstruct

(3) the Weak target image after strengthening is exported.