CN110472658A - A kind of the level fusion and extracting method of the detection of moving-target multi-source - Google Patents

Abstract

The invention belongs to the multi-source data levels based on Multiple Source Sensor to merge and extractive technique field, and in particular to a kind of the level fusion and extracting method of the detection of moving-target multi-source.The present invention will be seen that light image and infrared light image carry out registration fusion and obtain first layer blending image；After first layer blending image and high spectrum image are registrated, Weakening treatment is carried out to the image pixel after registration according to terrain classification region, obtains second layer blending image；Target acquisition is carried out to second layer blending image, the location information of target in the picture is obtained, target is perceived, obtain longitude and latitude of the target in true environment, the posture tracking target of aircraft is adjusted, realizes the lasting detection and perception to target.The present invention combines a variety of image sources, and the signal characteristic of a variety of image sources is effectively combined by image co-registration, removes the repeated data information of redundancy, increases the accuracy rate of target acquisition, improves detection efficient.

Description

A kind of the level fusion and extracting method of the detection of moving-target multi-source

Technical field

The invention belongs to the multi-source data levels based on Multiple Source Sensor to merge and extractive technique field, and in particular to a kind of The level fusion of moving-target multi-source detection and extracting method.

Background technique

With advances in technology with development, the quality of payload is obviously improved, and the increase of quality means More senser element can be carried, the computing capability and information storage capability of payload are also obviously improved, energy The calculating enough executed is more complicated.A variety of detecting devices can be often carried in the payload of spacecraft, it is such as visible Optical sensor, infrared sensor, bloom spectrum sensor etc. can obtain visible light sensing data image, infrared sensing data respectively Image, EO-1 hyperion sensing data image etc..

Premise for target tracking in image is target acquisition, how to establish a kind of quick, accurate and effective target spy Survey method is critical problem.Image recognition is that the methods and techniques of pattern-recognition are used in image domains.Pattern-recognition refers to pair Characterization things or the various forms of information of phenomenon are handled and are analyzed, things or phenomenon are described, recognized, is classified and The process of explanation is realized with computer and identifies and classify.By this thought in the picture, that is, realize similar to intelligence to sense The things known is recognized.The main thought of image recognition is the information bank for establishing the feature of things, is carried out to strange image Collected feature is compared collection apparatus with the information in known characteristic information library, to more than a certain similarity threshold The case where value, is both considered to find and identifies target.

Summary of the invention

Level fusion and extracting method the purpose of the present invention is to provide a kind of detection of moving-target multi-source.

The purpose of the present invention is realized by following technical solution: the following steps are included:

Step 1: reading the image of multi-source image sensor input；

Step 2: it will be seen that light image and infrared light image carry out image registration, the image after fusion registration obtains first Layer blending image；

Step 3: first layer blending image and high spectrum image are subjected to image registration, according to terrain classification region to registration after Image pixel carry out Weakening treatment, obtain second layer blending image；

Step 4: target acquisition being carried out to second layer blending image, the location information of target in the picture is obtained, to target It is perceived, obtains longitude and latitude of the target in true environment, adjust the posture tracking target of aircraft, target is held in realization Continuous detection and perception.

The present invention may also include:

Method for registering images in the step 2 and step 3 specifically:

Step 2.1: extracting image border profile, obtain the edge contour image of original image；

Profile, phase equalization function are extracted to image using phase equalization algorithm are as follows:

Wherein A_nFor the amplitude on scale n；φ_nIt (x) is n-th of Fourier components phase value at x；PC is worked as in expression (x) when obtaining maximum value at x, the weighted average amount of each component Local Phase parallactic angle of Fourier；

Step 2.2: the feature angle point with scale, position and direction information, specific side are established in edge contour image Method are as follows:

Step 2.2.1: construction Nonlinear Scale Space Theory makes feature angle point with dimensional information；

Gaussian filtering process is carried out to edge contour images, obtains image grey level histogram and contrast factor k；Conversion one Group uses additive operator splitting algorithm to obtain all Information Levels of nonlinear filtering image after calculating the time:

Wherein, A_lIndicate conductance matrix of the image I on different dimensions on l；t_iIt is defined as calculating the time, and only uses every time One group of calculating time constructs Nonlinear Scale Space Theory；E is unit battle array；

Step 2.2.2: detection feature angle point obtains characteristic angle dot position information；

A local window is moved point by point in the edge contour image of Nonlinear Scale Space Theory, to the pixel value in window Operation is carried out to determine whether being angle point；

Step 2.2.3: the directional information of feature angle point is calculated

The coordinate of feature angle point p (i) in the picture is (x (i), y (i)), and two point p (i-k), p (i+ are selected in neighborhood K), make the two points with point p (i) apart from being k, T is the tangent line at the place point p (i), and the principal direction of feature angle point p (i) is tangent line T and x The angle theta of axis positive direction_Feature, calculation formula is as follows:

Step 2.3: establishing shape description matrix；

If feature point set P { p₁,p₂,...p_n,},p_i∈R², with a certain characteristic point p (i) for origin, in being with p (i) point In r × r neighborhood of the heart, polar coordinate system is established, to 360 ° of 12 equal portions of progress, 12 sectors is obtained, is then followed successively by by radiusFive concentric circles are drawn, 60 zonules are obtained；The feature point number in each zonule is counted, is calculated p_iThe shape histogram hi of point, the shape histogram hi of each characteristic point are exactly Shape context description of each characteristic point； The calculation method of the shape histogram hi of each characteristic point are as follows:

Hi (k)=# { q ≠ p_i:(q-p_i)∈bin(k)}

Wherein # indicate statistics kth (k=1,2 ... 60) in a region characteristic point number；

Step 2.4: the characteristic angle point of two images being matched, image registration is completed；

The characteristic point of its arest neighbors and time neighbour, Euclidean distance are searched for by using Euclidean distance are as follows:

Wherein, a_iShape context for reference picture arbitrary characteristics point describes R (a₀,a₁,...a₅₉) in i-th, b_i Shape context for reference picture arbitrary characteristics point describes I (b₀,b₁,...b₅₉) in i-th；

If p is any feature point in certain piece image, with p arest neighbors to be registered and secondary neighbour's characteristic point be set to i, J, then the Euclidean distance of they and characteristic point p are respectively D_ipAnd D_jp；Set operation threshold valueAs the small Mr. Yu of the threshold value When one value, that is, thinks that p and i are the characteristic points correctly matched, otherwise fail.

The method of visible images and infrared light image after fusion is registrated in the step 2 specifically:

Step 3.1: region segmentation being carried out to the infrared image after registration, isolates suspicious region and the background of infrared image Region；The suspicious region is the highlight regions that the big image of infra-red radiation becomes clear；

Step 3.2: dual-tree complex wavelet transform being carried out with visible images to the infrared image after registration respectively, obtains image Low-frequency information and high-frequency information, the essential information of image correspond to the low-frequency information of wavelet transform result, the details letter of image Breath corresponds to the high-frequency information of wavelet transform result；

Step 3.3: the result of the result and wavelet transformation that divide the image into is merged, and respectively obtains low frequency blending image With high frequency blending image；

Step 3.4: low frequency blending image and high frequency blending image being subjected to dual-tree complex wavelet inverse transformation, obtain melting for the first time Close image.

Target acquisition is carried out to second layer blending image in the step 4, obtains the location information of target in the picture Method specifically:

Step 4.1: second layer blending image is filtered；

Establishing a window matrix, individual element scans on 2d, and the value of matrix center is in window matrix The average value of each point value substitutes, and indicates are as follows:

Wherein: f (x, y) is second layer blending image to be processed；G (x, y) is the second layer fusion figure after filtering processing Picture；S is the set with (x, y) point for middle neighborhood of a point coordinate points, and M is the sum for gathering internal coordinate；

Step 4.2: to the second layer blending image after filtering processing using the image threshold method of rolling average at Reason, obtains binary map；

Z_k+1Indicate the rolling average gray scale in scanning sequency, in the point that+1 step of kth encounters, at new point are as follows:

Wherein, n_{Gray scale}Represent the points used when calculating average gray, initial value m (1)=z_i/n_{Gray scale}；

Rolling average in the picture calculates every bit, so segmentation is executed with following formula:

Wherein, K is the constant in [0,1] range, m_xyIt is rolling average of the input picture at (x, y)；

Step 4.3: deleting the image that area is less than target in bianry image, remove the interference of irrelevant information；

Step 4.4: the bianry image after the interference of removal irrelevant information being handled using morphological image；

Step 4.5: establishing cutting function, target is cut into from the full image after morphological image process, is obtained To target image to be tested；

It is 0 that background parts, which are black level value, in image I after morphological image process, and target part to be tested is white value It is 1；From image (0,0), coordinate starts to have looked for, and finds the point that first coordinate pixel value is 1, since looking for this point, finds All the points are named as set T by the point that all pixels value being attached thereto is 1₁, in set T₁In point coordinate in find cross The maximum value x of coordinate_1maxWith minimum value x_1min, in set T₁In point coordinate in find the maximum value y of ordinate_1maxMost Small value y_1min, then the target image to be tested that cutsAnd so on, all targets to be tested are found, are obtained To all target images to be tested

Step 4.6: finding the main symmetry axis of target image to be tested using principal component analytical method, and obtain target image to be tested Main symmetry axis and x-axis angle theta_{It is to be tested}；

The coordinate of each point is two dimension in target image information to be tested, these are put composition n_{It is to be tested}2 column matrix X of row_{It is to be tested}, Middle n_{It is to be tested}For the points in target image information to be tested, X is calculated_{It is to be tested}Covariance matrix C_{It is to be tested}, and continue to calculate covariance square Battle array C_{It is to be tested}Feature vector V_{It is to be tested}=(x_v,y_v), then the main symmetry axis of target image to be tested and x-axis angle theta_{It is to be tested}Are as follows:

Step 4.7: carrying out image direction normalization, target image to be tested is rotated into θ_{It is to be tested}Angle, and remove newly generated black Side；

Step 4.8: image size normalization is carried out, it will be by the target image picture size to be tested of direction normalization Become template size；

Step 4.9: by the target image to be tested after direction normalizing and size normalization and the image in template library It is matched one by one, and sets similar threshold value T, be target by this image recognition when similarity degree is more than this threshold value.

The method of low frequency blending image is obtained in the step 3.3 are as follows: according to infrared Image Segmentation at suspicious region and The location information of background area, it will be seen that light image is split according to identical location information；For infrared image with it is visible The suspicious region of light image low frequency part, using following rule:

Wherein,For l layers of fusion figure low frequency coefficient,For l layers of infrared figure Picture low frequency coefficient,For l layers of visible images low frequency coefficient；

For the background area of infrared image and visible images low frequency part, using Local Deviation method, Local Deviation is got over Greatly, show that how corresponding each pixel gray-value variation in the region be bigger, the pixel contrast in the opposite region is just Higher, the corresponding information in this region is also more；The pixel big to Local Deviation value increases weight in image co-registration, rule It is then as follows:

Wherein, ω_irFor infrared image weight, ω_visFor visible images weight；Infrared image weight ω_irAnd visible light Image weights ω_visCalculation method are as follows:

ω_ir=1- ω_vis；

Wherein, σ_visAnd σ_irThe respectively Local Deviation of visible images and infrared image, r are related coefficient region；It can be seen that The Local Deviation σ of light image_visWith the Local Deviation σ of infrared image_irCalculation method are as follows:

The calculation method of correlation coefficient r are as follows:

Wherein image size M × N,Indicate the average gray value of visible images,Indicate the average ash of infrared image Angle value, I_ir(i, j) represents infrared image, I_vis(i, j) represents visible images.

The beneficial effects of the present invention are:

The present invention applies visible images, infrared light image and high spectrum image simultaneously, while the visible images point that get both Resolution is high, and infrared light image target contrast is high, and high-spectrum can be as distinguishing the various sensor maps such as artificiality and natural object The advantages of picture, it is accurate that target acquisition determines, influence of the earth atmosphere activity to target acquisition is effectively reduced.The present invention combines a variety of Image source, by image co-registration can effectively combine a variety of image sources signal characteristic, effectively remove the repeat number of redundancy it is believed that Breath improves detection efficient so as to effectively increase the accuracy rate of target acquisition.The tracking to target may be implemented in the present invention, And the prediction of the location informations such as bogey heading, the speed of a ship or plane, longitude and latitude.

Detailed description of the invention

Fig. 1 is overall flow schematic diagram of the invention.

Fig. 2 is that infrared image of the invention merges flow diagram with visible light image registration.

Fig. 3 is that first layer blending image of the invention merges flow diagram with high spectrum image.

Fig. 4 is target acquisition and perception flow diagram of the invention.

Specific embodiment

The present invention is described further with reference to the accompanying drawing.

The invention belongs to based on Multiple Source Sensor multi-source data level fusion with extractive technique field, in spacecraft Payload in can often carry a variety of detecting devices, such as visible light sensor, infrared sensor, bloom spectrum sensor Deng can obtain visible light sensing data image, infrared sensing data image, EO-1 hyperion sensing data image etc. respectively, specifically relate to And multi-source image fusion, to technologies such as the perception of the identification of image, target

Image data is originated from multiple sensors in the present invention, and often same position target will appear multiple images, still The information emphasis that image from every kind of sensor carries is different, and due to similar background information between different images Data redundancy is caused, so before executing the image processing techniques such as image segmentation, target acquisition, target acquisition, target apperception, Fusion treatment is carried out to image, bulk information data are subjected to integration screening, to remove redundancy, leave different data sources Valid data information in image.Image fusion technology includes image registration techniques and fusing image data technology.

The purpose of the present invention is utilizing multi-source image sensor, consider attitude of flight vehicle is motor-driven, spatial attitude disturbance, cloud Caused image fault under the factors such as platform vibration；Consider to cause under the conditions ofs earth atmosphere motion artifacts, sea sea situation are severe etc. Image object masking, sufficiently with different sensors to the sharp sex differernce of the different characteristic of signal, pass through the image of multilayer Registration and image co-registration make image reach image definition requirements, the requirement of target standout, and reache a certain level and show The requirement for the target that is blocked, based on fused image, design object probe algorithm reaches in multiregion, more images The target acquisition in source, and combine aircraft appearance rail information, attitude of flight vehicle motor-driven etc., reach in certain time space to target Discovery, tracking and position prediction function.

Purpose of the present invention implementation:

Step 1: reading the image of multi-source image sensor input；

Step 2: it will be seen that light image and infrared light image carry out image registration, the image after fusion registration obtains first Layer blending image；

Step 3: first layer blending image and high spectrum image are subjected to image registration, according to terrain classification region to registration after Image pixel carry out Weakening treatment, obtain second layer blending image；

Step 4: target acquisition being carried out to second layer blending image, the location information of target in the picture is obtained, to target It is perceived, obtains longitude and latitude of the target in true environment, adjust the posture tracking target of aircraft, target is held in realization Continuous detection and perception.

1) first layer merges: establishing infrared image and visible light image registration and blending algorithm, utilizes visible light resolution ratio The features such as height, but vulnerable to atmospheric effect, infrared image target contrast is high, but environment resolution ratio is low, make fused image mesh Mark is prominent and ambient enviroment has certain details.

2) second layer merges: that establishes high spectrum image and first layer blending image is registrated blending algorithm, reaches prominent mesh Mark weakens the purpose that irrelevant information influences.

3) image recognition is carried out to second layer blending image.It establishes a kind of with certain actual application ability and feasibility Target acquisition algorithm.

4) target apperception, on the basis of 3) step discovery target, in conjunction under aircraft appearance rail information, earth rotation, star Information, the rough spotting earth longitudes and latitudes such as point position vector, optical axis direction vector predict target on the basis of target trajectory Track is perceived discovery.

By above step, the independent discovery of payload, identification, the ability for tracking target can be reached.

Advantages of the present invention mainly has a two o'clock: first, image source of the present invention has the whole world, round-the-clock on aircraft, The features such as range is wide, high-timeliness.Second, the present invention applies visible images, infrared light image and high spectrum image simultaneously, together When get both visible images high resolution, infrared light image target contrast is high, high-spectrum can as distinguish artificiality with from The advantages of various sensor images such as right object, sufficiently improve target acquisition rate and accuracy rate.Third, the present invention make full use of platform Advantage may be implemented to a degree of tracking of target and the prediction of position (course, the speed of a ship or plane, longitude and latitude).

Firstly, the problems such as considering due to aircraft platforms vibration and focusing deviation is so infrared light image and visible light figure As image registration problem must be taken into consideration before fusion.Consider that objective contour is core in infrared image and visible images, so choosing With the Data Matching fusion method based on shape descriptor.Quickly similar nearest neighbor algorithm is selected in image registration of the invention, is examined Worry wants existing this algorithm widely to be used in the image of such situation, stability and universality with higher.Herein On the basis of, RANSAC algorithm is selected in order to further enhance the robustness of inventive algorithm, to matched characteristic point It is screened, rejects erroneous point, be left optimal collection.

Considering high-altitude image imaging characteristics, imaging is influenced vulnerable to earth atmosphere activity, and considers spacecraft orbit limitation, at The problems such as forming an angle with plumb line as optical axis, cause scalloping, image region segmentation method of the invention is to be based on What saliency was considered, Image Fusion is also based on the result of segmentation.Consider the imaging characteristics of infrared image, long wave Infrared image is to acquire object itself to exhale the long-wave infrared come, and the infra-red radiation of target area is often higher than environment , thus we consider to enhance this graphic processing method using conspicuousness, by infrared image target area in image procossing Enhancing, background information weaken, and play the role of inhibiting noise to a certain extent.It can be efficiently separated in conjunction with dual-tree complex wavelet algorithm The characteristic of high-frequency information in image and low-frequency information out considers infrared image and can be by the imaging characteristics of light figure and image matter Amount uses different convergence strategies to high frequency low frequency respectively.Complete first layer image co-registration.

Secondly, high spectrum image has irreplaceable advantage in the research of analysis atural object, the present invention uses frequency spectrum phase Like degree classification method, to improve the accurate fixed of ground object target detection, on the basis of distinguishing atural object is inartificial or artificiality, then A tomographic image fusion is carried out, achievees the purpose that weaken the jamming patterns information such as the inhuman divine force that created the universe, be provided effectively for final discrimination objective Foundation and effectively support.Complete second layer image co-registration.

Finally, being merged based on final image, it is contemplated that the redundancy in image is effectively weakened, and target highlights, The present invention includes a kind of image recognition algorithm based on template matching, and can persisted moniker target position, track, and certain Target direction (such as longitude and latitude, course, the speed of a ship or plane) is predicted in degree, attitude of flight vehicle is instructed to change, and is guaranteed to Target self-determination Lasting identification and tracking.

Heterologous sensing data used in the present invention is respectively from visible light sensor and infrared sensor.Both sensings Feature of image obtained by device is as shown in table 1 below.Load is on aircraft platforms, when the image of acquisition, it is contemplated that ground Ball air motion, Ambient reflect scattering process；And it in view of aircraft platforms pose adjustment, platform disturbance vibration and puts down Platform stability etc.；And in view of photoelectric platform assembling and setting, platform property etc., picture position, side caused by many factors Position, scale, shape difference.So this method is intended to reduce influence of the extraneous factor to image, and maximum journey to the greatest extent The advantage of two kinds of sensor acquisition data of protrusion of degree.

1 visible light sensor of table and infrared sensor acquire feature of image

The present invention includes process of image registration.First using phase equalization algorithm (Phase Congruency) to image Extract profile.Phase equalization function is defined as:

Wherein A_nFor the amplitude on scale n；φ_nIt (x) is n-th of Fourier components phase value at x；PC is worked as in expression (x) when obtaining maximum value at x, the weighted average amount of each component Local Phase parallactic angle of Fourier.Obtain the edge contour of original image Image is applied in subsequent processing.

The present invention includes to establish feature point methods.Nonlinear Scale Space Theory building method is as follows, and input picture is carried out Then gaussian filtering process obtains image grey level histogram, obtain contrast factor k.Reconvert uses after one group of calculating time Additive operator splitting algorithm obtains all Information Levels of nonlinear filtering image:

Wherein, A_lIndicate conductance matrix of the image I on different dimensions on l.t_iIt is defined as calculating the time, and only uses every time One group of calculating time constructs Nonlinear Scale Space Theory, and E is unit battle array.

Angular-point detection method is as follows, moves a local window point by point in the picture, transports to the pixel value in window It calculates to determine whether being angle point.Assuming that the grey scale change that local window C translation (u, v) generates afterwards, it may be assumed that

Wherein I (x, y) is gray value of the image in (x, y) point, and w (x, y) indicates gaussian weighing function.

In order to acquire the value for keeping E (u, v) as big as possible, above formula Taylor expansion obtains:

And then switch to matrix form:

WhereinI_x,I_yDivide in the x-direction with the gradient in the direction y for image grayscale Amount.

Then local autocorrelation function E (u, v) can be approximated to be elliptic function:

E(u,v)≈Au²+2Cuv+Bv²

Equal reference points around the point constitute elliptic curve, and the point on ellipse is identical as central point degree of correlation, second order The eigenvalue λ of matrix M₁,λ₂It is elliptical long axis and short axle respectively, indicates the direction of the speed of grey scale change.So working as characteristic value λ₁,λ₂It is angle point when all very big and suitable.When two characteristic values small one and large one when, be edge.When two characteristic values all very littles When, as flat site.In order to make image that there is the scale invariability in Corner Detection, Corner Detection Algorithm is brought into Into Nonlinear Scale Space Theory described above, makes characteristic point simultaneous with scale and location information, obtain angle point receptance function Are as follows:

Wherein, σ_i,SFor scale factor,Respectively the second-order differential of grey scale change in the x and y direction and partially Derivative.The point for meeting angle point receptance function is angle point.

Angle point is characterized plus suitable directional information.The coordinate of known features angle point p (i) in the picture is (x (i), y (i)) two point p (i-k), p (i+k), are selected in neighborhood, make the two points with point p (i) apart from being k, T is the place point p (i) Tangent line, the principal direction of feature angle point p (i) are the angle theta of tangent line T and positive direction of the x-axis_Feature, calculation formula is as follows:

Angle point is characterized on the basis of above and determines principal direction, the advantage of doing so is that making it have rotational invariance. It can be good at being used in the feature angle point matching problem between infrared image and visible images, hereinafter referred to as characteristic point.

The present invention includes shape descriptor generating algorithm.If feature point set P { p₁,p₂,...p_n,},p_i∈R², with a certain spy Levying point p (i) is origin, in r × r neighborhood centered on p (i) point, establishes polar coordinate system, to 360 ° of 12 equal portions of progress, obtains To 12 sectors, then it is followed successively by by radiusFive concentric circles are drawn, to obtain 60 zonules. The feature point number in each zonule is counted, p is calculated_iThe shape histogram hi of point, hi are defined as follows:

Hi (k)=# { q ≠ p_i:(q-p_i)∈bin(k)}

Wherein # indicate statistics kth (k=1,2 ... 60) in a region characteristic point number.The shape of each characteristic point is straight Side's figure is exactly Shape context description of each characteristic point.

The present invention includes the matching process of feature point set.Arest neighbors and time neighbour's characteristic point are searched for, most using quick approximation Nearest neighbor algorithm searches for the characteristic point that each characteristic point searches for its arest neighbors and time neighbour by using Euclidean distance.Euclidean distance is fixed Justice are as follows:

Wherein, a_iShape context for reference picture arbitrary characteristics point describes R (a₀,a₁,...a₅₉) in i-th, b_i Shape context for reference picture arbitrary characteristics point describes I (b₀,b₁,...b₅₉) in i-th.Then algorithm concrete operations walk It suddenly is, if p is any feature point in infrared image, to be set to i, j with p arest neighbors to be registered and time neighbour's characteristic point, then The Euclidean distance of they and characteristic point p are respectively D_ipAnd D_jp.Set operation threshold valueWhen the threshold value is less than a certain value When, that is, think that p and i are the characteristic points correctly matched, otherwise fails.

RANSAC algorithm is selected in order to further enhance the robustness of inventive algorithm, to matched characteristic point It is screened, rejects erroneous point, be left optimal collection.All best match characteristic points are substituted into image sky to location parameter by the algorithm Between projective transformation model, image projection transformation relationship is obtained by direct linear transformation's algorithm, the registration parameter of image is exactly red Affine transformation relationship between outer image and visible images.So far, we complete matching between infrared image and visible images Quasi- process.

The present invention includes image co-registration process.Its process is summarized as first carrying out region segmentation to infrared image, can separate The highlight regions of infrared image and background area out, then using the result of Infrared Image Region Segmentation Based as foundation, to visible images Make corresponding mapping.Dual-tree complex wavelet transform carried out respectively to infrared image and visible images, processing the result is that To the low-frequency information and high-frequency information of image, the essential information of image corresponds to the low-frequency information of wavelet transform result, image Detailed information corresponds to the high-frequency information of wavelet transform result.The result coupling of the result and wavelet transformation that divide the image into is examined Consider.When handling low-frequency information, highlight regions and background area, difference and the mission requirements of the information reflected are considered Consider, reply actual conditions use different convergence strategies.When handling high-frequency information, it is contemplated that high-frequency information is mainly schemed anyway The minutia of picture is melted so not being all that weight is distributed in each region according to the abundant degree of detailed information according to weight design Close strategy.The specific steps of above-mentioned process are as follows:

Step 3.1: region segmentation being carried out to the infrared image after registration, isolates suspicious region and the background of infrared image Region；The suspicious region is the highlight regions that the big image of infra-red radiation becomes clear；

Step 3.2: dual-tree complex wavelet transform being carried out with visible images to the infrared image after registration respectively, obtains image Low-frequency information and high-frequency information, the essential information of image correspond to the low-frequency information of wavelet transform result, the details letter of image Breath corresponds to the high-frequency information of wavelet transform result；

Step 3.3: the result of the result and wavelet transformation that divide the image into is merged, and respectively obtains low frequency blending image With high frequency blending image.

The highlight regions of infrared image are chosen, the image for carrying out conspicuousness enhancing to the infrared image after registration is first had to Processing, the thermal target information enhancement of infrared image after processing, background information is fuzzy, so that the contrast of entire infrared image increases By force.The algorithm of conspicuousness enhancing is mainly based upon the histogram of image.Pixel I in image I_cConspicuousness is defined as:

Wherein Dis (I_c,I_i)=| | I_c-I_i| | it is I_cColor distance indicates I_cWith I_iDifference in color, above formula can change It is written as:

Wherein a_cFor pixel I_cGray value, n be image in contained gray scale sum, f_jFor a_jWhat is occurred in the picture is general Rate；Image significance figure is calculated, saliency map I is obtained_sal；

Inclusion region partitioning algorithm of the present invention.Each pixel in image is mixed into spy by K gauss hybrid models Levying indicates, wherein K=k ∈ { 1,2 ... K }.Pixel one kind in image is corresponding with target gauss hybrid models, another It is corresponding with background gauss hybrid models.The image Gibbs energy function of Region Segmentation Algorithm as a result, are as follows:

E (α, k, θ, z)=U (α, k, θ, z)+V (α, z)

Wherein z is pixel value, and corresponding pixel belongs to background, α corresponding pixel when being 1 when α ∈ { 0,1 }, α are 0 Belong to target, U is area item, and V is border item, the calculation method of area item U are as follows:

θ={ π (α, k), μ (α, k), ∑ (α, k), α=0,1, k=1...K }

Area item is used to distinguish the pixel in target area or the pixel in background area, determines parameter θ by learning Afterwards, the region energy item of Gibbs determines that.

The calculation method of border item V are as follows:

Wherein γ is empirical, is obtained by training, C indicates the set of adjacent pixel pair, function [α_n≠α_m] value Only 1 or 0, work as α_n≠α_mWhen, [α_n≠α_m]=1, works as α_n=α_mWhen, [α_n≠α_m]=0.

β=(2 < | | z_m-z_n||²>)^-1, indicate the mathematic expectaion of sample.The contrast that β corresponds to image can determine right The border item of the situation more high or low than degree.Image is split with max-flow min-cut algorithm, after the completion of segmentation, optimization gauss Mixed model parameter, repeatedly iteration completes image segmentation when making energy function minimum.This image partition method is transported It need to be by saliency map I mentioned above when using the present invention_salAs the initialization value of image segmentation algorithm, it is used to uncalibrated image Highlight regions and background area be iterated segmentation according to the region demarcated, obtain segmentation result.

The present invention include based on dual-tree complex wavelet transform (dual-tree complex wavelet transform, DTCWT Image Fusion Rule), dual-tree complex wavelet function is defined as:

ψ (x)=ψ_h(x)+jψ_g(x)

Wherein ψ_h(x) and ψ_gIt (x) is real number small echo.After two-dimentional DTCWT transformation, picture breakdown obtains two low frequency wavelets The high frequency coefficient of coefficient and six direction (± 15 °, ± 45 °, ± 75 °).

For low frequency part first by region segmentation method mentioned above, infrared image is split respectively, point Cut out suspicious region and except exterior domain, record position information, it will be seen that light image is split according to identical location information, if Resolution ratio difference then, then after normalizing location information, is handled as coefficient.

According to infrared Image Segmentation at the location information of suspicious region and background area, it will be seen that light image is according to identical Location information is split；For the suspicious region of infrared image and visible images low frequency part, using following rule:

Wherein,For l layers of fusion figure low frequency coefficient,For l layers of infrared figure Picture low frequency coefficient,For l layers of visible images low frequency coefficient；

Exterior domain is removed for low frequency part, using Local Deviation method, Local Deviation is bigger, shows each in the region How corresponding pixel gray-value variation be bigger, and the pixel contrast in the opposite region is higher, one can consider that this The corresponding information in region is also more.According to the above analysis, the pixel big to Local Deviation value increases power in image co-registration Weight, rule are as follows:

For the background area of infrared image and visible images low frequency part, using Local Deviation method, Local Deviation is got over Greatly, show that how corresponding each pixel gray-value variation in the region be bigger, the pixel contrast in the opposite region is just It is higher, one can consider that the corresponding information in this region is also more.According to the above analysis pixel big to Local Deviation value Point increases weight in image co-registration, and rule is as follows:

Wherein, ω_irFor infrared image weight, ω_visFor visible images weight；Infrared image weight ω_irAnd visible light Image weights ω_visCalculation method are as follows:

ω_ir=1- ω_vis；

Wherein, σ_visAnd σ_irThe respectively Local Deviation of visible images and infrared image, r are related coefficient region；It can be seen that The Local Deviation σ of light image_visWith the Local Deviation σ of infrared image_irCalculation method are as follows:

The calculation method of correlation coefficient r are as follows:

Wherein image size M × N,Indicate the average gray value of visible images,Indicate the average ash of infrared image Angle value, I_ir(i, j) represents infrared image, I_vis(i, j) represents visible images.

High frequency section is divided the image by image partition method above as n region, with A={ a₁, a₂...a_nIndicate.If each region corresponds to the region weight of oneselfThe present invention provides following weighting High frequency fusion rule.

Wherein, a numerical value C is set^l,θ> 1 is used to amplify high frequency coefficient, it is therefore an objective to pair of detail section in prominent image Degree of ratio.It is contemplated that doing so, the noise in image can equally be amplified, so a two values matrix M is added^l,θ, whenWhen M^l,θ=1, and going divider value is 1 but isolated point, and doing so only amplification to connect sheet of high frequency coefficient picture Vegetarian refreshments removes noise；For contracting function, it is therefore an objective to reduce influence of the noise to high-frequency information.During actual fused The concave-convex variation at edge may cause the result of fusion to be distorted deformation, we are using infrared image and visible images The high frequency coefficient that DTWCT is converted calculates a unit vector, improves, is then merged at this time to original high frequency coefficient The high frequency coefficient of image is rewritten are as follows:

Wherein:Wherein, it merges Regular f is with infrared image and can be by weight in light imageOn the basis of big, the region r of the image is calculated_iHigh frequency coefficient Mean value, then using this mean value as the correspondence high frequency coefficient of blending image.S^l,θBy highlight regions image segmentation result elder generation dilation operation It is obtained using 2 dimension mean filters, it is therefore an objective to guarantee that blending image each cell domain detailed information is more significant.Wherein region is weighed Value

Wherein H^l,θ(x, y) high frequency coefficient, l are the number of plies, and θ is directional subband, | r_i ^θ| it is region r_i ^θSize.

The present invention includes that high spectrum image active layer time characters of ground object extraction and image co-registration become owner of image method, such as attached drawing 3 It is shown.

Firstly, it is generally believed that high-spectrum remote-sensing refers to spectral resolution 10^-2The remote sensing of λ order of magnitude range.High-spectrum The features such as feature of picture, wave band is more, and spectral region is narrow, and wave spectrum is continuous, it is even hundreds of that the single pixel member in image contains tens A wave band, the range of each wave band are less than the spectrum of 10nm.So analyzing from spectral Dimensions remote sensing information, analysis is different Information bank is demarcated and established to the feature of object reflective spectrum, and knowledge is matched in information bank using the spectroscopic data of target Not, to add label in image object, the identification to atural object is realized with this.

Atural object differentiation is carried out from frequency domain angle, is believed the corresponding complete spectrum of each pixel of Hyperspectral imaging as sequence Number, classified using frequency spectrum similarity classification method (FSSM method) to trial zone image.Due to high-spectral data have from Type is dissipated, is analyzed so discrete Fourier transform (DFT) can be used, DFT transform can with compressed signal and largely Upper inhibition noise and Hughes phenomenon, and obtain signal spectrum, this makes it possible to effectively extract differently, object light spectral curve is not The main wave crest of co-wavelength position and the frequency spectrum of trough retain the effective information on the curve of spectrum.

One-dimensional discrete Fourier transformation is used first, and spectral signal is transformed into frequency domain and obtains frequency spectrum.It will be every in HIS image The corresponding spectral sequence of a pixel is considered as one-dimensional discrete signal f (n), and DFT may be defined as:

WhereinP (k)=R²(k)+I²(k)；

F_phase=arctan (I (k)/R (k))；

In formula | F (k) |, P (k), F_phaseIt is amplitude spectrum, energy spectrum and the phase of the corresponding spectral sequence of pixel respectively Spectrum, R (k), I (k) are the real and imaginary parts of F (k) respectively, and k is the serial number of DFT transform, and N is discrete sampling data length, n be from Scattered sampled point, that is, corresponding high-spectral data wave band number, f (n) are pixels in the corresponding reflectance value of each wave band, i.e. ground light Compose reflectance value.

Next yin calculates the difference of frequency spectrum between target spectrum and reference spectra frequency spectrum with lance distance, and then measures spectrum Similitude to classify, calculation formula is as follows:

In formula, F_tar(i) and F_ref(i) be respectively target and reference spectra curve frequency spectrum, N_sIt is the low order for participating in calculating Multi harmonics.Reference spectra can be using the pixel spectrum extracted on Laboratory Spectra, field measuring spectrum or image.It uses When the spectrum of field measuring is as reference spectra, in order to eliminate influence of the atmosphere to spectrum, need to carry out atmosphere to remote sensing image It corrects.

Consider the region split in first layer fusion.The simple and effective of consideration method, i.e., reduce as far as possible Requirement of the algorithm to hardware computing resource, storage resources.Consider material and water, rock, the plant of nature etc. of man-made target 's.And just for the background of so-called target, natural object is all exactly background.So only needing to identify in high spectrum image The natural background part come, is weakened in the first tomographic image by the method for weighting and is attenuated to the influence of non-targeted information It is minimum.Specific method for registering images is the same as the method in first layer fusion.

The present invention include moving-target detection and tracing algorithm, as shown in Fig. 4.

Target acquisition function:

Final image is filtered first, purpose removes noise.Mean filter process is to establish a window square Individual element scans battle array on 2d, and the average value of each point value substitutes in the value window matrix of matrix center, It is represented by

Wherein: f (x, y) is second layer blending image to be processed；G (x, y) is the second layer fusion figure after filtering processing Picture；S is the set with (x, y) point for middle neighborhood of a point coordinate points, and M is the sum for gathering internal coordinate；

Then it is handled using the image threshold method of rolling average, grayscale image is made to become binary map.Basic thought is Along the scanning line computation rolling average of piece image.With the execution of zigzag mode by-line, and then reduce illumination deviation.Enable Z_k+1 It indicates in scanning sequency, in the point that+1 step of kth encounters.Rolling average gray scale at new point is given by:

Wherein, n_{Gray scale}Represent the points used when calculating average gray, initial value m (1)=z_i/n_{Gray scale}；

Rolling average in the picture calculates every bit, so segmentation is executed with following formula:

Wherein, K is the constant in [0,1] range, m_xyIt is rolling average of the input picture at (x, y)；

It is 5 times of target width, K=0.5 that usually we, which take n,.This Research on threshold selection can effectively avoid shadow point Target image is extracted in the influence that cloth unevenness generates the variation of image two-value, help.

Consider that the single width of hardware computing capability and permission handles the time, scaling appropriate can be carried out to image in terms of improving Calculate speed.Only consider the biggish image down processing of length not being related to magnification distortion problem, so in selecting method herein When only need to consider simplicity, therefore herein the present invention select arest neighbors difference arithmetic

Then cell area image unrelated with target in image is deleted, area is deleted in bianry image and is significantly less than target Image, to delete the interference of irrelevant information in image.

Image is handled using image aspects mathematic(al) function, makes to obtain shape the most essential after image object is processed Feature.Two basic morphology operations are understood first.Erosion operation is defined as:

Wherein A ⊙ S expression corrodes A using S, and detailed process is that structural element S is allowed to move in the plane domain of A image Dynamic, if S can be completely contained in A when the origin translation of S to z point, the set that all such z points are constituted is denoted as S To the corrosion image of A.Corrosion can melt the boundary of object, can make thin connection position fracture in image object.

Dilation operation is defined as:

WhereinExpression expands A using S, and detailed process is to allow structural element S in the whole image plane of A The interior movement mapping of S relative to its own origin when its own origin translation to z pointThere is common intersection with A, i.e.,Extremely with A Rare 1 pixel is overlapping, then the set that all such z points are constituted is denoted as S to the expanding image of A.Expansion energy makes object Boundary expands, and can connect the gap of fracture.

Opening operation is to carry out dilation operation after first carrying out erosion operation to image A using structural element S, is indicated are as follows:

Opening operation can eliminate that small object, separating objects, smooth big object boundary do not change volume at very thin point.

Closed operation is to carry out erosion operation after first carrying out dilation operation to image A using structural element S, is indicated are as follows:

Closed operation being capable of filler body inner aperture hole, connection adjacent object, unobvious object area and the shape of changing In the case of its smooth boundary.

The present invention can select corresponding Morphology Algorithm according to the actual situation, reach and eventually find suspected target.

Cutting function is established, target is cut into from full image, obtains target image to be tested.Method is, according to preceding Text processing as a result, it is 0 that background parts, which are black level value, in image I, target part to be tested is that white value is 1.It is sat from image (0,0) Mark starts to have looked for, and finds the point that first coordinate pixel value is 1, since looking for this point, finds all pictures being attached thereto All the points are named as set T by the point that element value is 1₁, in set T₁In point coordinate in find the maximum value x of abscissa_1max With minimum value x_1min, in set T₁In point coordinate in find the maximum value y of ordinate_1maxWith minimum value y_1min, then cut The target image to be tested arrivedx_1min< x < x_1max,y_1min< y < y_1max.And so on, it finds all to be tested Target obtains all target images to be tested

In view of the man-made target that we are identified usually has symmetry, so we are looked for using principal component analytical method To the main symmetry axis of image to be tested, and obtain itself and x-axis angle theta.Principal component analytical method is the maximum that it is found out in N-dimensional data Distribution vector and value.The coordinate of each point is two dimension in target image information to be tested, these are put composition n_{It is to be tested}2 column matrix of row X_{It is to be tested}, wherein n_{It is to be tested}For the points in target image information to be tested, X is calculated_{It is to be tested}Covariance matrix C_{It is to be tested}, and continue to calculate association Variance matrix C_{It is to be tested}Feature vector V_{It is to be tested}=(x_v,y_v), then the main symmetry axis of target image to be tested and x-axis angle theta_{It is to be tested}Are as follows:

Then image direction normalization is carried out, by image rotation θ_{It is to be tested}Angle, and newly generated black surround is removed, both cut again It cuts.

Then image size normalization is carried out, picture size is become into template size.Template library is established, and provides template Having a size of M × N in library.

Establish target discrimination function:

By treated, image to be tested carries out matching one by one with the image in template library and sets a certain similar threshold value T.When It is target by this image recognition when similarity degree is more than this threshold value.Specific matching step is as follows:

1) similarity threshold T is set

2) H is calculated according to the following method, it is assumed that image A to be tested, template image B.Judge image pixel value A (x, y) to be tested It is whether equal with template image pixel value B (x, y).The H=H+1 if equal；Next point is judged if unequal, wherein x ∈ (0, M),y∈(0,N).Obtain final H value

If 3) judge H > T, it is judged as discovery target；If H < T, next template is replaced, step 2) is repeated, until All templates are gone through, then is judged as and does not find target.

Target apperception:

Application target probe function obtains target position, i.e., coordinate in the picture.

(1) according to load position, target position longitude and latitude is determined.

(2) position coordinates according to target in the picture determine pose adjustment angle, and target is maintained at image central area In domain.

The above two o'clock reaches target apperception purpose.

The foregoing is only a preferred embodiment of the present invention, is not intended to restrict the invention, for the skill of this field For art personnel, the invention may be variously modified and varied.All within the spirits and principles of the present invention, made any to repair Change, equivalent replacement, improvement etc., should all be included in the protection scope of the present invention.

Claims (7)

1. the level fusion and extracting method of a kind of moving-target multi-source detection, it is characterised in that: the following steps are included:

Step 1: reading the image of multi-source image sensor input；

Step 2: it will be seen that light image and infrared light image carry out image registration, the image after fusion registration obtains first layer and melts Close image；

Step 3: first layer blending image and high spectrum image being subjected to image registration, according to terrain classification region to the figure after registration As pixel progress Weakening treatment, second layer blending image is obtained；

Step 4: target acquisition being carried out to second layer blending image, obtains the location information of target in the picture, target is carried out Perception, obtains longitude and latitude of the target in true environment, adjusts the posture tracking target of aircraft, realizes the lasting spy to target It surveys and perceives.

2. the level fusion and extracting method of a kind of moving-target multi-source detection according to claim 1, it is characterised in that: institute Method for registering images in the step 2 and step 3 stated specifically:

Step 2.1: extracting image border profile, obtain the edge contour image of original image；

Profile, phase equalization function are extracted to image using phase equalization algorithm are as follows:

Wherein A_nFor the amplitude on scale n；φ_nIt (x) is n-th of Fourier components phase value at x；It indicates to work as PC (x) In When obtaining maximum value at x, the weighted average amount of each component Local Phase parallactic angle of Fourier；

Step 2.2: the feature angle point with scale, position and direction information is established in edge contour image, method particularly includes:

Step 2.2.1: construction Nonlinear Scale Space Theory makes feature angle point with dimensional information；

Gaussian filtering process is carried out to edge contour images, obtains image grey level histogram and contrast factor k；Convert one group of meter All Information Levels of nonlinear filtering image are obtained using additive operator splitting algorithm after evaluation time:

Wherein, A_lIndicate conductance matrix of the image I on different dimensions on l；t_iIt is defined as calculating the time, and every time only with one group The time is calculated to construct Nonlinear Scale Space Theory；E is unit battle array；

Step 2.2.2: detection feature angle point obtains characteristic angle dot position information；

It moves a local window point by point in the edge contour image of Nonlinear Scale Space Theory, the pixel value in window is carried out Operation is to determine whether be angle point；

Step 2.2.3: the directional information of feature angle point is calculated

The coordinate of feature angle point p (i) in the picture is (x (i), y (i)), and two point p (i-k), p (i+k) are selected in neighborhood, Make the two points with point p (i) apart from being k, T is the tangent line at the place point p (i), and the principal direction of feature angle point p (i) is tangent line T and x-axis The angle theta of positive direction_Feature, calculation formula is as follows:

Step 2.3: establishing shape description matrix；

If feature point set P { p₁,p₂,...p_n,},p_i∈R², with a certain characteristic point p (i) for origin, in the r centered on p (i) point In × r neighborhood, polar coordinate system is established, to 360 ° of 12 equal portions of progress, 12 sectors is obtained, is then followed successively by by radiusFive concentric circles are drawn, 60 zonules are obtained；The feature point number in each zonule is counted, is calculated p_iThe shape histogram hi of point, the shape histogram hi of each characteristic point are exactly Shape context description of each characteristic point； The calculation method of the shape histogram hi of each characteristic point are as follows:

Hi (k)=# { q ≠ p_i:(q-p_i)∈bin(k)}

Wherein # indicate statistics kth (k=1,2 ... 60) in a region characteristic point number；

Step 2.4: the characteristic angle point of two images being matched, image registration is completed；

The characteristic point of its arest neighbors and time neighbour, Euclidean distance are searched for by using Euclidean distance are as follows:

Wherein, a_iShape context for reference picture arbitrary characteristics point describes R (a₀,a₁,...a₅₉) in i-th, b_iFor ginseng The Shape context for examining image arbitrary characteristics point describes I (b₀,b₁,...b₅₉) in i-th；

If p is any feature point in certain piece image, it is set to i, j with p arest neighbors to be registered and time neighbour's characteristic point, then The Euclidean distance of they and characteristic point p are respectively D_ipAnd D_jp；Set operation threshold valueWhen the threshold value is less than a certain value When, that is, think that p and i are the characteristic points correctly matched, otherwise fails.

3. a kind of level fusion of moving-target multi-source detection according to claim 1 or 2 and extracting method, feature exist In: the method for visible images and infrared light image after fusion is registrated in the step 2 specifically:

Step 3.1: region segmentation being carried out to the infrared image after registration, isolates suspicious region and the background area of infrared image Domain；The suspicious region is the highlight regions that the big image of infra-red radiation becomes clear；

Step 3.2: dual-tree complex wavelet transform being carried out with visible images to the infrared image after registration respectively, obtains the low of image Frequency information and high-frequency information, the essential information of image correspond to the low-frequency information of wavelet transform result, the detailed information pair of image It should be in the high-frequency information of wavelet transform result；

Step 3.3: the result of the result and wavelet transformation that divide the image into is merged, and respectively obtains low frequency blending image and height Frequency blending image；

Step 3.4: low frequency blending image and high frequency blending image being subjected to dual-tree complex wavelet inverse transformation, obtain first time fusion figure Picture.

4. a kind of level fusion of moving-target multi-source detection according to claim 1 or 2 and extracting method, feature exist In: target acquisition, the method for obtaining the location information of target in the picture are carried out to second layer blending image in the step 4 Specifically:

Step 4.1: second layer blending image is filtered；

Establishing a window matrix, individual element scans on 2d, each point in the value window matrix of matrix center The average value of value substitutes, and indicates are as follows:

Wherein: f (x, y) is second layer blending image to be processed；G (x, y) is the second layer blending image after filtering processing；S It is the set with (x, y) point for middle neighborhood of a point coordinate points, M is the sum for gathering internal coordinate；

Step 4.2: the second layer blending image after filtering processing being handled using the image threshold method of rolling average, is obtained To binary map；

Z_k+1Indicate the rolling average gray scale in scanning sequency, in the point that+1 step of kth encounters, at new point are as follows:

Wherein, n_{Gray scale}Represent the points used when calculating average gray, initial value m (1)=z_i/n_{Gray scale}；

Rolling average in the picture calculates every bit, so segmentation is executed with following formula:

Wherein, K is the constant in [0,1] range, m_xyIt is rolling average of the input picture at (x, y)；

Step 4.3: deleting the image that area is less than target in bianry image, remove the interference of irrelevant information；

Step 4.4: the bianry image after the interference of removal irrelevant information being handled using morphological image；

Step 4.5: establish cutting function, target be cut into from the full image after morphological image process, obtain to Test target image；

It is 0 that background parts, which are black level value, in image I after morphological image process, and target part to be tested is that white value is 1； From image (0,0), coordinate starts to have looked for, and finds the point that first coordinate pixel value is 1, since looking for this point, finds therewith All the points are named as set T by the point that connected all pixels value is 1₁, in set T₁In point coordinate in find abscissa Maximum value x_1maxWith minimum value x_1min, in set T₁In point coordinate in find the maximum value y of ordinate_1maxAnd minimum value y_1min, then the target image to be tested that cutsx_1min< x < x_1max,y_1min< y < y_1max, and so on, All targets to be tested are found, all target images to be tested are obtained

Step 4.6: finding the main symmetry axis of target image to be tested using principal component analytical method, and obtain target image master couple to be tested Claim axis and x-axis angle theta_{It is to be tested}；

The coordinate of each point is two dimension in target image information to be tested, these are put composition n_{It is to be tested}2 column matrix X of row_{It is to be tested}, wherein n_{It is to be tested}For the points in target image information to be tested, X is calculated_{It is to be tested}Covariance matrix C_{It is to be tested}, and continue to calculate covariance matrix C_{It is to be tested}Feature vector V_{It is to be tested}=(x_v,y_v), then the main symmetry axis of target image to be tested and x-axis angle theta_{It is to be tested}Are as follows:

Step 4.7: carrying out image direction normalization, target image to be tested is rotated into θ_{It is to be tested}Angle, and remove newly generated black surround；

Step 4.8: carrying out image size normalization, will become by the target image picture size to be tested of direction normalization Template size；

Step 4.9: one by one by the target image to be tested after direction normalizing and size normalization and the image in template library It is matched, and sets similar threshold value T, be target by this image recognition when similarity degree is more than this threshold value.

5. the level fusion and extracting method of a kind of moving-target multi-source detection according to claim 3, it is characterised in that: institute Target acquisition is carried out to second layer blending image in the step 4 stated, the method for obtaining the location information of target in the picture is specific Are as follows:

Step 4.1: second layer blending image is filtered；

Establishing a window matrix, individual element scans on 2d, each point in the value window matrix of matrix center The average value of value substitutes, and indicates are as follows:

Wherein: f (x, y) is second layer blending image to be processed；G (x, y) is the second layer blending image after filtering processing；S It is the set with (x, y) point for middle neighborhood of a point coordinate points, M is the sum for gathering internal coordinate；

Step 4.2: the second layer blending image after filtering processing being handled using the image threshold method of rolling average, is obtained To binary map；

Z_k+1Indicate the rolling average gray scale in scanning sequency, in the point that+1 step of kth encounters, at new point are as follows:

Wherein, n_{Gray scale}Represent the points used when calculating average gray, initial value m (1)=z_i/n_{Gray scale}；

Rolling average in the picture calculates every bit, so segmentation is executed with following formula:

Wherein, K is the constant in [0,1] range, m_xyIt is rolling average of the input picture at (x, y)；

Step 4.3: deleting the image that area is less than target in bianry image, remove the interference of irrelevant information；

Step 4.4: the bianry image after the interference of removal irrelevant information being handled using morphological image；

Step 4.5: establish cutting function, target be cut into from the full image after morphological image process, obtain to Test target image；

It is 0 that background parts, which are black level value, in image I after morphological image process, and target part to be tested is that white value is 1； From image (0,0), coordinate starts to have looked for, and finds the point that first coordinate pixel value is 1, since looking for this point, finds therewith All the points are named as set T by the point that connected all pixels value is 1₁, in set T₁In point coordinate in find abscissa Maximum value x_1maxWith minimum value x_1min, in set T₁In point coordinate in find the maximum value y of ordinate_1maxAnd minimum value y_1min, then the target image to be tested that cutsx_1min< x < x_1max,y_1min< y < y_1max, and so on, All targets to be tested are found, all target images to be tested are obtained

Step 4.6: finding the main symmetry axis of target image to be tested using principal component analytical method, and obtain target image master couple to be tested Claim axis and x-axis angle theta_{It is to be tested}；

The coordinate of each point is two dimension in target image information to be tested, these are put composition n_{It is to be tested}2 column matrix X of row_{It is to be tested}, wherein n_{It is to be tested}For the points in target image information to be tested, X is calculated_{It is to be tested}Covariance matrix C_{It is to be tested}, and continue to calculate covariance matrix C_{It is to be tested}Feature vector V_{It is to be tested}=(x_v,y_v), then the main symmetry axis of target image to be tested and x-axis angle theta_{It is to be tested}Are as follows:

Step 4.7: carrying out image direction normalization, target image to be tested is rotated into θ_{It is to be tested}Angle, and remove newly generated black surround；

Step 4.8: carrying out image size normalization, will become by the target image picture size to be tested of direction normalization Template size；

Step 4.9: one by one by the target image to be tested after direction normalizing and size normalization and the image in template library It is matched, and sets similar threshold value T, be target by this image recognition when similarity degree is more than this threshold value.

6. the level fusion and extracting method of a kind of moving-target multi-source detection according to claim 3, it is characterised in that: institute The method of low frequency blending image is obtained in the step 3.3 stated are as follows: according to infrared Image Segmentation at suspicious region and background area Location information, it will be seen that light image is split according to identical location information；For infrared image and visible images low frequency Partial suspicious region, using following rule:

Wherein,For l layers of fusion figure low frequency coefficient,For l layers of infrared image low frequency Coefficient,For l layers of visible images low frequency coefficient；

For the background area of infrared image and visible images low frequency part, using Local Deviation method, Local Deviation is bigger, table How corresponding each pixel gray-value variation in the bright region be bigger, and the pixel contrast in the opposite region is higher, The corresponding information in this region is also more；The pixel big to Local Deviation value increases weight in image co-registration, and rule is such as Under:

Wherein, ω_irFor infrared image weight, ω_visFor visible images weight；Infrared image weight ω_irIt is weighed with visible images Value ω_visCalculation method are as follows:

ω_ir=1- ω_vis；

Wherein, σ_visAnd σ_irThe respectively Local Deviation of visible images and infrared image, r are related coefficient region；Visible light figure The Local Deviation σ of picture_visWith the Local Deviation σ of infrared image_irCalculation method are as follows:

The calculation method of correlation coefficient r are as follows:

Wherein image size M × N,Indicate the average gray value of visible images,Indicate the average gray of infrared image Value, I_ir(i, j) represents infrared image, I_vis(i, j) represents visible images.

7. the level fusion and extracting method of a kind of moving-target multi-source detection according to claim 5, it is characterised in that: institute The method of low frequency blending image is obtained in the step 3.3 stated are as follows: according to infrared Image Segmentation at suspicious region and background area Location information, it will be seen that light image is split according to identical location information；For infrared image and visible images low frequency Partial suspicious region, using following rule:

Wherein,For l layers of fusion figure low frequency coefficient,For l layers of infrared image low frequency Coefficient,For l layers of visible images low frequency coefficient；

For the background area of infrared image and visible images low frequency part, using Local Deviation method, Local Deviation is bigger, table How corresponding each pixel gray-value variation in the bright region be bigger, and the pixel contrast in the opposite region is higher, The corresponding information in this region is also more；The pixel big to Local Deviation value increases weight in image co-registration, and rule is such as Under:

Wherein, ω_irFor infrared image weight, ω_visFor visible images weight；Infrared image weight ω_irIt is weighed with visible images Value ω_visCalculation method are as follows:

ω_ir=1- ω_vis；

Wherein, σ_visAnd σ_irThe respectively Local Deviation of visible images and infrared image, r are related coefficient region；Visible light figure The Local Deviation σ of picture_visWith the Local Deviation σ of infrared image_irCalculation method are as follows:

The calculation method of correlation coefficient r are as follows:

Wherein image size M × N,Indicate the average gray value of visible images,Indicate the average gray of infrared image Value, I_ir(i, j) represents infrared image, I_vis(i, j) represents visible images.