CN101295401A - Infrared point target detecting method based on linear PCA - Google Patents

Abstract

The invention relates to a detection method based on a linear PCA for infrared point targets, comprising the following steps of: first, creating a training sample of the infrared point target through an improved gaussian gray model, and carrying out trainings for the PCA through the samples, to produce a group of principal components; then intercepting and acquiring sub-images from an infrared image to be detected, and respectively projecting the sub-images to each principal component, to reconstruct the sub-images through obtained projection coefficients and the principal components, and calculate the reconstruction error; finally, substituting a measure function with the reconstruction error, to produce a detection image and improve the testability of the target. The detection method based on the linear PCA for infrared point targets can greatly improve the testability of the infrared point targets. Different from the filtration-based detection method, the detection method based on the linear PCA for infrared point targets does not require a preprocessing for the infrared images to be detected.

Description

Infrared point target detecting method based on linear PCA

Technical field

The present invention relates to a kind of infrared target detection method of technical field of image processing, specifically is a kind of infrared point target detecting method based on linear PCA.

Background technology

In video monitoring and search system, the infrared photography The Application of Technology has very big singularity, can all weather operations, and be a kind of passive passive detection technology, can be hidden well when finding target oneself.Wherein, how detecting target under remote condition, is one of gordian technique of infrared monitoring and search system.If can detect target on farther distance, system just has the time of more target being reacted, for the system status that has the initiative in antagonism creates conditions.In addition, target detection is the first step that system reacts to target, and the performance of target detection directly has influence on the effect of subsequent treatment (such as target following and Target Recognition).Therefore, the infrared target detection technique is for the infrared video system, and is significant.

When the target range ir imaging system was far away, the area of target on imager was very little, is the little target of point-like (usually all less than 100 pixels).Because the influence of factors such as the contrast of infrared point target is low, area is little, the interference of texture and edge fog, noise and clutter is stronger makes the detection of infrared point target very difficult.Traditional infrared point target detecting method great majority are based on wave filter, have based on the method for airspace filter device: the filtering of high pass template, medium filtering, mathematical morphology filter, the filtering of local standard difference etc.; Method based on frequency domain filter has: desirable high-pass filtering, Butterworth high-pass filtering, Gauss's high-pass filtering method etc.If view data (in the image be that the gray-scale value of all pixels in the subimage at center arrange the vector that form with each pixel) is considered as the data set is made up of target class data and background classes data, the problem of detection target just converts to target data is discerned the problem that (classification) comes out from view data from image.Therefore, some mode identification methods can be used for realizing target detection.PCA (principal component analysis (PCA)) is an important method in the pattern recognition theory, often is used to realize data compression.Obtain describing the proper vector (principal component, major component) of training data by training, promptly Zui Da several characteristic is worth pairing proper vector.PCA has also extracted the feature of training data when realizing data compression.

Find by prior art documents, China's application number: 200410068024.7, patent name is " morphologic filter automatic destination detecting method ", this patent detects target based on the morphologic filtering method of Top-hat operator, concrete principle is: little target is in high band, and background is in low-frequency range, like this by High frequency filter or deduct the image of low-pass filtering gained with original image, the purpose that can realize giving prominence to target and suppress background, thus target detection finished.Obviously, this method is when detecting target, and some noise spots (also being in high band) also have bigger output.Like this, noise spot can be a target by flase drop also, and has improved false alarm rate, and this is based on the inherent shortcoming of the infrared point target detecting method of wave filter.

Summary of the invention

Of the present inventionly propose a kind of infrared point target detecting method, can realize detection the infrared point target of various wave bands based on linear PCA in order to solve the above-mentioned problems in the prior art.

The present invention is achieved by the following technical solutions, the present invention includes following steps:

At first,, PCA is trained, produce one group of principal component by these samples with the training sample of improved Gauss's gray level model generation infrared point target;

Then, from infrared image to be detected, intercept subimage, and it is projected to respectively on each principal component, this subimage is reconstructed, and calculate reconstructed error with the projection coefficient and the principal component that obtain;

At last,, produce detected image, improved the detectability of target reconstructed error substitution detection function.

Described improvement Gauss gray level model produces the infrared point target training sample, is specially: at traditional Gauss's gray level model $I(i,j)=I_{\max }\exp (-\frac{1}{2}[\frac{{(i-x_0)}^2}{{\mathrm{\σ}}_x^2}+\frac{{(j-y_0)}^2}{{\mathrm{\σ}}_y^2}])$ The basis on, add two constraint conditions $T_1\≤\frac{I_{\max }}{{\mathrm{\σ}}_x}\≤T_2$ With $T_1\≤\frac{I_{\max }}{{\mathrm{\σ}}_y}\≤T_2,$ Make it to become improved Gauss's gray level model:

$I(i,j)=I_{\max }\exp (-\frac{1}{2}[\frac{{(i-x_0)}^2}{{\mathrm{\σ}}_x^2}+\frac{{(j-y_0)}^2}{{\mathrm{\σ}}_y^2}]),$

$s.t.\left\{\begin{array}{c}{T}_1\≤\frac{I_{\max }}{{\mathrm{\σ}}_x}\≤T_2\\ T_1\≤\frac{I_{\max }}{{\mathrm{\σ}}_y}\≤T_2\end{array}\right..$

So just finished the improvement of Gauss's gray level model.Wherein, I _MaxBe target's center's pixel value (gray scale peak value); σ _xBe horizontal dispersion parameter, σ _yBe the vertical dispersion parameter, controlling the stroll characteristic of object pixel; (x ₀, y ₀) be the centre coordinate of target image; (i j) is other pixel coordinate of target image.I _Max, σ _xAnd σ _yThree parameters have been determined the characteristic of Gauss's gray level model.The infrared point target that traditional Gauss's gray level model produces can produce some and " point target " can not occur in actual infrared image, (works as I such as producing some point targets that are similar to noise spot _MaxBigger, and σ _xAnd σ _yHour), if train PCA, can reduce its detection performance greatly with such training sample (sample exterior point Outlier).

Describedly with sample PCA is trained, produce one group of principal component, be specially: produce some infrared point target images with above-mentioned improved Gauss's gray level model, these images are as training sample.With the grey scale pixel value of the image pattern of each infrared point target of producing, joining by the row first place is arranged in vector, with these vectors PCA is trained, and the generation principal component is got n proper vector of character pair value maximum, finishes target detection.The n value detects effect by compromise consideration and computation complexity (amplitude of dimensionality reduction) is decided, and is preferably n=6.

Describedly intercept subimage from infrared image to be detected, and it is projected to respectively on each principal component, with the projection coefficient and the principal component that obtain this subimage is reconstructed, and calculate reconstructed error, concrete steps are:

Step 1, (x, y) position intercepting subimage P in detected image _s, and generate vectorial P _Sc, with P _ScProject to respectively on n the principal component, obtain n projection coordinate, formula is

Wherein, ω _kBe P _ScK proper vector

On projection coordinate, Γ is an average criterion;

Step 2 is reconstructed subimage with n principal component and n projection coordinate, obtains reconstructed image, and formula is

Wherein,

Be reconstructed image;

Step 3, the reconstructed error of calculating subimage, formula is

Wherein, ε _RCBe reconstructed error.If subimage P _sBe target image, so reconstructed image

With P _sSimilarity just very big, otherwise then similarity is very little, so reconstructed error be exactly this subimage whether be that a kind of of target image estimates, the smaller value correspondence of reconstructed error the existence of target.

Described with reconstructed error substitution detection function, produce detected image, be specially: with reconstructed error ε _RCThe substitution detection function ${\mathrm{\ζ}}_D(x,y)=e^{-{\mathrm{\ϵ}}_{\mathrm{RC}}{(x,y)}^2/2{\mathrm{\δ}}^2}.$ Different with reconstructed error is ζ _D(x, higher value correspondence y) the existence of target, and this is consistent with human intuition; And ζ _D(x, value y) has saved the normalization link between 0～1, be more convenient for calculating detection indexs such as signal to noise ratio (snr).The value of parameter δ wherein is chosen for δ=0.25 by experience.With detection function value ζ corresponding on each location of pixels _D(x y), as the gray-scale value of this pixel, has just generated detected image.

Traditional infrared point target detection great majority are based on filtering method, when detecting target, also the noise spot erroneous judgement are impact point, have improved false alarm rate, have increased the load of system.The present invention is that the method with pattern-recognition detects infrared point target, PCA is an important branch of pattern-recognition, often be used to realize data compression, but in packed data, PCA has also extracted clarification of objective, utilize this characteristic, PCA is used to realize the detection of infrared point target, its essence is with PCA in detected image point target " identification " to be come out.For recognition methods, training sample is very big to the influence of its discrimination, and the present invention at first improves Gauss's gray level model, produces the training sample of infrared point target then with improved Gauss's gray level model.Through PCA is trained, produce principal component, detected subimage (intercepting from detected image) is projected to respectively on each principal component, obtain the respective projection coordinate, by the reconstruct of projection coordinate and principal component to subimage, calculate reconstructed error, less reconstructed error correspondence the existence of target.A detection function is devised, and is used for producing detected image, and in detected image, the detectability of infrared point target is improved greatly.

The present invention can improve the detectability of infrared point target largely, and different with method based on filtering, this method does not need infrared image to be detected is carried out pre-service.

Description of drawings

Fig. 1 is the training sample by the infrared point target of improved Gauss's gray level model generation

Fig. 2 is the synoptic diagram of the embodiment of the invention

Wherein: (a₁)、(b ₁) be respectively and comprise two original infrared images of 10 point targets, figure (b₁) in target 6 is a real Ship Target, and other target in two width of cloth images all is to be produced by improved Gauss's gray level model Simulation objectives embed and to obtain in the infrared image, all targets among the figure all are Weak targets, detection difficulty is bigger; (a₂)、(b ₂) be respectively PCA to (a₁)、(b ₁) result that detects; (a₃)、(b ₃) be respectively (a₂)、(b ₂) 3D Structure chart.

Embodiment

Below in conjunction with specific embodiment technical scheme of the present invention is elaborated: present embodiment is being to implement under the prerequisite with the technical solution of the present invention; provided detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to following embodiment.

Present embodiment has mainly comprised the content of two aspects, i.e. the production method of infrared point target training sample and based on the infrared point target detecting method of PCA.

As shown in Figure 1, be the training sample of the infrared point target that produced by improved Gauss's gray level model, wherein: the size of training sample is 11 * 11, and the image that is produced by improved Gauss's gray level model rotates at any angle and obtains.

(1) training

1) grey scale pixel value of 121 target training sample subimages (11 * 11) is arranged by end to end order, generated column vector (121 dimension);

2) column vector is arranged generator matrix, and matrix is carried out characteristic value decomposition, obtain optimum description clarification of objective vector, and proper vector is arranged from big to small by eigenwert; Present embodiment is got preceding 6 proper vectors

These 6 proper vectors are to describe an eigenvectors of target.

(2) detect

3) each location of pixels in detected image is used the sliding window with target sample picture size size identical (11 * 11), intercepts subimage to be detected, and converts thereof into column vector (121 dimension);

4) column vector that subimage to be detected generated is projected on 6 proper vectors, correspondingly obtain 6 projection coordinate { ω ₁, ω ₂... ω _n;

5) with these 6 proper vectors and projection coordinate subimage to be detected is reconstructed, and calculates reconstructed error ε _RC

6) with ε _RCSubstitution detection function ζ _D(x y), calculates the detection function value on each location of pixels, with the gray-scale value of these functional values as respective pixel, generates detected image.

Fig. 2 is the testing result of embodiment, (a among the figure ₂), (b ₂) be respectively PCA to (a ₁), (b ₁) result that detects; (a ₃), (b ₃) be respectively (a ₂), (b ₂) the 3D structural drawing.Table 1 has provided the detection index of Fig. 2, is respectively snr gain (SNRG) and background inhibiting factor (BSF).Snr gain is the tolerance to the signal to noise ratio (S/N ratio) load-carrying capacity, and computing formula is SNRG=SNR _Out/ SNR _In, SNR wherein _OutBe the signal to noise ratio (S/N ratio) of detected image ((a2) among Fig. 2 and (b2)), SNR _InSignal to noise ratio (S/N ratio) for original image ((a1) among Fig. 2 and (b1)); The background inhibiting factor is background to be suppressed the tolerance of ability, and computing formula is BSF=C _In/ C _Out, C wherein _InBe the standard deviation of original image background, C _OutStandard deviation for the detected image background.Two detection indexs are big more, and the detection performance of detection method is good more.Can find out that from table 1 a few target in Fig. 2 (b1), the snr gain of the inventive method gained illustrates that all greater than 1 this method can improve the detectability of infrared target.In addition, to two width of cloth images of embodiment, the background inhibiting factor illustrates that all much larger than 1 this method has very strong background to suppress ability.

Table 1 three seed space arithmetics compare the performance index of target detection in the accompanying drawing 2

Claims (5)

1. the infrared point target detecting method based on linear PCA is characterized in that, comprises the steps:

At first,, PCA is trained, produce one group of principal component by these samples with the training sample of improved Gauss's gray level model generation infrared point target;

Then, from infrared image to be detected, intercept subimage, and it is projected to respectively on each principal component, this subimage is reconstructed, and calculate reconstructed error with the projection coefficient and the principal component that obtain;

At last, with reconstructed error substitution detection function, produce detected image;

Described improved Gauss's gray level model is specially:

$I(i,j)=I_{\max }\exp (-\frac{1}{2}[\frac{{(i-x_0)}^2}{{\mathrm{\σ}}_x^2}+\frac{{(j-y_0)}^2}{{\mathrm{\σ}}_y^2}\left]\right),$

$s.t.\left\{\begin{array}{c}{T}_1\≤\frac{I_{\max }}{{\mathrm{\σ}}_x}\≤T_2\\ T_1\≤\frac{I_{\max }}{{\mathrm{\σ}}_y}\≤T_2\end{array}\right..$

Wherein, I _MaxIt is target's center's pixel value; σ _xBe horizontal dispersion parameter, σ _yBe the vertical dispersion parameter, controlling the stroll characteristic of object pixel; (x ₀, y ₀) be the centre coordinate of target image; (i j) is other pixel coordinate of target image, I _Max, σ _xAnd σ _yThree parameters have been determined the characteristic of Gauss's gray level model.

2. the infrared point target detecting method based on linear PCA according to claim 1, it is characterized in that, describedly PCA is trained with sample, produce one group of principal component, be specially: produce some infrared point target images with improved Gauss's gray level model, these images are as training sample, grey scale pixel value with the image pattern of each infrared point target of producing, join and be arranged in vector by the row first place, with these vectors PCA is trained, produce principal component, get n proper vector of character pair value maximum, finish target detection.

3. the infrared point target detecting method based on linear PCA according to claim 2 is characterized in that, described n value is 6.

4. the infrared point target detecting method based on linear PCA according to claim 1, it is characterized in that, describedly from infrared image to be detected, intercept subimage, and it is projected to respectively on each principal component, with projection coefficient that obtains and principal component this subimage is reconstructed, and the calculating reconstructed error, concrete steps are:

Step 1, (x, y) position intercepting subimage P in detected image _s, and generate vectorial P _Sc, with P _ScProject to respectively on n the principal component, obtain n projection coordinate, formula is

Wherein, ω _kBe P _ScK proper vector On projection coordinate, Γ is an average criterion;

Step 2 is reconstructed subimage with n principal component and n projection coordinate, obtains reconstructed image, and formula is

Wherein, Be reconstructed image;

Step 3, the reconstructed error of calculating subimage, formula is Wherein, ε _RCBe reconstructed error, if subimage P _sBe target image, so reconstructed image With P _sSimilarity just very big, otherwise then similarity is very little, so reconstructed error be exactly this subimage whether be that a kind of of target image estimates.

5. the infrared point target detecting method based on linear PCA according to claim 1 is characterized in that, and is described with reconstructed error substitution detection function, produces detected image, is meant: with reconstructed error ε _RCThe substitution detection function ${\mathrm{\ζ}}_D(x,y)=e^{-{\mathrm{\ϵ}}_{\mathrm{RC}}{(x,y)}^2/2{\mathrm{\δ}}^2},$ The value of parameter δ wherein is taken as δ=0.25, with detection function value ζ corresponding on each location of pixels _D(x y), as the gray-scale value of this pixel, has just generated detected image.

Images