CN102201058B - Cat eye effect object recognition algorithm of active and passive imaging system sharing same aperture - Google Patents

Abstract

The invention provides a cat eye effect object recognition algorithm of active and passive imaging systems sharing a same aperture. The algorithm is suitable for detecting and recognizing optical targets with a characteristic of cat eye effect. The algorithm of the invention improves the cat eye effect object recognition algorithm based on gray scales and shape features. For the active and passive imaging systems sharing the same aperture, differential operation is performed between active and passive images after image registration so that background can be filtered. Rough registering algorithm can be improved. A search scope of matching feature points in a second image can be reduced in horizontal and vertical directions. And algorithm efficiency can be raised. The algorithm possesses good utility value and wide application prospect in the fields of laser imaging and digital image processing technology.

Description

" opal " effect Target Recognition Algorithms of active imaging and passive imaging of common aperture system

(1) technical field

The present invention is for the Target Recognition Algorithms of taking aim at equipment with optoelectronic device or the sight of optical lens.Utilize active imaging and passive imaging of common aperture system acquisition master by motion video and identification " opal " effect target, belong to laser imaging and digital image processing techniques fields.

(2) background technology

when one of laser alignment during with the optical system of optical lens, optical system can be similar to the combination of regarding lens and a detector focal plane as, laser in its visual field is through the effect of optical system, form one " image patch " near the focal plane of optical system, this " image patch " is equivalent to a light source, laser reflects back with certain beam divergence angle through optical system, by the light path principle of reversibility, laser beam can be returned along former road, thereby generation good directionality, the reflected in parallel light of concentration of energy, its light intensity substantially exceeds common diffuse reflection light intensity, " opal " effect of Here it is imaging optical system.

" opal " performance characteristic of optical system has very high using value, based on the Laser Active Detection system of " opal " effect, because of its advantage with early warning, is widely used.The active imaging and passive imaging of common aperture system is a kind of of Laser Active Detection system.At present, the LASER Light Source of the Laser Active Detection system of " opal " effect has continuous laser source and two kinds of forms of pulsed laser light source." opal " effect active detection system of pulsed laser light source does not have good dynamic scan characteristic, and the dynamic scan characteristic of continuous laser source " opal " effect active detection system is good, can be used in the The Cloud Terrace system to carry out servo-actuated search sweep; Pulsed laser light source " opal " effect active detection system need to gather video flowing to Same Scene and carry out subsequent treatment, this algorithm uses continuous laser source " opal " effect active detection system only need gather main passive two width images to Same Scene and gets final product (initiatively image namely relies on from the main light source emergent light and reflects formed picture through target to be measured, is namely relied on target own radiation light or reflects the picture that natural light forms by motion video).Continuous laser source " opal " effect active detection system is adopted in domestic identification to " opal " effect target more, and Same Scene only gathers initiatively image of a width, and " opal " effect target is determined in the recognition methods of based target gray scale and shape facility.Yet the method target recognition effect in complex background is not good, lacks reliability, and this patent algorithm improves this algorithm, leads passive image difference computing after having added image registration, uses at last shape to differentiate locking " opal " effect target.

The active imaging and passive imaging of common aperture system receives light and at first (adopts Cold Mirrors, i.e. transmitted infrared light through the dichroism spectroscope as shown in Figure 1, reflect visible light) be divided into two-way, transmitted light is through CCD1 (Charge-coupled Device, i.e. charge-coupled image sensor) imaging, for initiatively looking like; Reflected light is passive picture through the CCD2 imaging.

(3) summary of the invention

1, purpose: the recognition methods that the present invention is directed to intensity-based and shape facility is unsuitable for the defective of target identification in complex background, proposes the improvement algorithm for common aperture imaging system, namely carries out the image difference computing after image registration.The original image that the method is processed be continuous laser as lighting source gather main by motion video pair, utilize main passive image difference computing to realize the background filtering, and this algorithm has added image registration techniques before using main passive image difference computing, greatly reduces because main passive image pixel subtraction point does not represent the interference noise that the same object point of scene produces.Use this algorithm identified " opal " effect target precise and high efficiency more.

2, technical scheme: the present invention by two CCD cameras gather respectively main, by motion video, the image of at first one of them CCD camera being taken carries out horizontal mirror image switch to be processed, then image is carried out gaussian filtering, then carries out image registration.Image registration comprises feature point detection and Feature Points Matching two parts, wherein, at first the coupling of unique point adopts the Euclidean distance of proper vector as similarity measurement, obtain the feature point set of thick coupling, adopt again the RANSAC (abbreviation of RANdom SAmpleConsensus of high robust, be the random sampling unification algorism) algorithm iteration purification matching double points data, and calculate the transformation relation of two width images in conjunction with least square method, obtain perspective transformation matrix.Utilize the perspective transformation matrix that obtains in the Feature Points Matching process, piece image is mapped on the matched position of another width image, then carry out the image difference computing.Differentiate locking " opal " effect target by shape at last.

Fig. 2 is the process flow diagram of this algorithm, and the specific implementation step is as follows:

Step 1: image level mirror image switch and gaussian filtering;

Step 2: image registration;

Step 3: image conversion;

Step 4: image difference;

Step 5: image binaryzation;

Step 6: zone marker;

Step 7: shape facility is differentiated;

Step 8: target area locking.

3, advantage and effect: advantage of the present invention is to lead passive image difference computing after adopting image registration again, realizes the background filtering.Because two CCD put error, mainly, subtracted each other pixel and do not necessarily represent same object point in scene during without the computing of images match direct differential by motion video, this algorithm has overcome this drawback, has improved the accuracy of identification " opal " effect target.

(4) description of drawings

Fig. 1 is active imaging and passive imaging of common aperture system architecture schematic diagram.This system architecture is mainly by continuous semiconductor laser instrument (CWLASER), Laser emission and receiving optics, dichroism spectroscope, two compositions such as CCD camera, DSP (DigitalSignal Processing, digital signal processing) processing unit and display module.Lasing light emitter is made of semiconductor laser, exports continuous infrared laser; The optical axis of CCD1, CCD2 camera lens orthogonal and all with dichroism spectroscope angle at 45 °; After the processing unit processes conversion of image through digital signal processor DSP, shown by display module.

Fig. 2 is algorithm flow chart.

(5) embodiment

The present invention adopts continuous laser as active light source, can realize round-the-clock target identification, uses the dichroism spectroscope to see through the infrared light reflection visible light, uses two CCD cameras to obtain respectively main by motion video.

Fig. 2 is algorithm flow chart of the present invention, and these algorithm concrete steps are as follows:

1, image level mirror image switch and gaussian filtering: obtain the main by motion video of Same Scene by two CCD, system architecture determines that CCD1 and CCD2 are imaged as horizontal mirror, need to carry out horizontal mirror image switch to piece image wherein, re-use Gaussian filter to master, passive image filtering;

2, image registration: image registration comprises feature point detection and Feature Points Matching two parts.Use SIFT (Scale-InvariantFeature Transform, i.e. yardstick invariant features conversion) algorithm to carry out feature point detection, to image I ₁, I ₂Detect respectively N ₁, N ₂Individual unique point.The SIFT algorithm is a kind of algorithm that extracts local feature, search out extreme point at metric space, extracting position, yardstick, invariable rotary characteristic quantity are therefore the SIFT feature descriptor has good robustness to illumination variation, image rotation, proportional zoom, geometry deformation, fuzzy and compression of images.

At first the coupling of unique point adopts the Euclidean distance of proper vector as similarity measurement, obtains the feature point set of thick coupling.Get image I ₁In certain unique point, in image I ₂Middle searching and its Euclidean distance recently and time two near unique points, if minimum distance thinks that except in proper order closely less than certain threshold value this is a pair of match point.There is the mistake coupling unavoidably in above thick matching result, then adopts RANSAC algorithm iteration purification match point.The RANSAC algorithm is a kind of robust Model algorithm for estimating of widespread use, main thought is: the as far as possible little sampling set of initial utilization, estimate model parameter, then under estimated parameter, select the data point of data centralization to enlarge initial sampling set as far as possible, iteration produces the most homogeneous data set, utilizes at last the most homogeneous data set, re-starts the estimation of model parameter.The model parameter of determining at last consists of perspective transition matrix H, and the unique point that consistent data is concentrated is to satisfying:

$\left(\begin{array}{c}{x}_i\\ y_i\\ 1\end{array}\right)=H*\left(\begin{array}{c}{x}_i^{\′}\\ y_i^{\′}\\ 1\end{array}\right)$

(x wherein _i, y _i) be I ₁In i (1≤i≤N ₁) individual unique point coordinate, (x ' _i, y ' _i) be image I ₂In characteristic of correspondence point coordinate with it.Do not think that at the matching double points of most homogeneous data centralization the point of matching error is right.

This patent improves thick matching algorithm in process of image registration.To main used SIFT feature point detection algorithm by motion video after, image I ₁, I ₂N is arranged respectively ₁, N ₂Individual unique point judges image I in former thick matching process ₁In i unique point whether with image I ₂In certain Feature Points Matching the time, computed image I ₁In i the point and I ₂All N ₂The Euclidean distance of individual unique point and sequence.When calculating I ₁In all N ₁During the matching relationship of individual unique point, need repetitive operation N ₁Inferior, work as N ₁And N ₂When all very large, calculate very consuming timely, efficiency of algorithm is low.For the active imaging and passive imaging of common aperture system, by to the images after registration computational analysis, knowing that two width image corresponding point row pixels and row pixel coordinate do not differ can surpass 50.If therefore I ₁The pixel coordinate of middle unique point i is (a, b), at I ₂During the corresponding point of middle search i, the hunting zone can be restricted to, row-coordinate at a-50 between a+50, the row coordinate at b-50 to the unique point between b+50.

3, image conversion: utilize the perspective transformation matrix that obtains in the Feature Points Matching process, piece image is mapped on the matched position of another width image.

$\left[\begin{array}{c}{x}^{\′}\\ y^{\′}\\ 1\end{array}\right]~H*\left[\begin{array}{c}x\\ y\\ 1\end{array}\right]$

What formula represented is with image I ₂Be mapped to image I ₁Matched position on, [x y 1] is I ₂The homogeneous coordinates of middle pixel, [x ' y ' 1] is I ₂The homogeneous coordinates of the pixel after projective transformation.

4, image difference: with image I ₁With image I ₂Image after projective transformation carries out the image difference computing;

5, image binaryzation: use self-adaption thresholding method, image after calculus of differences is carried out binary conversion treatment;

6, zone marker: have a plurality of connected domains in binary image, for extracting its feature, need to carry out zone marker;

7, shape facility is differentiated: pass through comprehensive criterion again

Metric＝|1-Metric _r|+|1-Metric _e|

Further lock onto target region.Wherein

Represent the circularity of cut zone,

Represent the excentricity of cut zone, A is the area of cut zone, and P is the girth of cut zone, and a is the major axis of cut zone, and b is the minor axis of cut zone.The more approaching zero explanation critical region of the value of Metric is near circle;

8, lock onto target zone: realized the target area locking by above flow process.

Advantage of the present invention is to introduce image registration techniques to lead passive image difference computing and realize the background filtering, has improved the reliability of " opal " effect target identification.In process of image registration, thick matching algorithm is improved, for the active imaging and passive imaging of common aperture system, dwindled from the ranks both direction hunting zone of seeking unique point to be matched the second width image, improved efficiency of algorithm.

Claims (2)

1. " opal " effect Target Recognition Algorithms of active imaging and passive imaging of common aperture system, lead passive image difference computing after the employing image registration techniques and realize the background filtering, improved the reliability of " opal " effect target identification, can identify accurately " opal " effect target in complex background, algorithm flow is summarised as following key step:

(1) image of CCD in system being taken carries out horizontal mirror image switch to be processed, then to the master, carried out gaussian filtering by motion video;

(2) carry out image registration after filtering, removed because main passive image pixel subtraction point does not represent the interference noise that in scene, same object point brings;

Use the SIFT algorithm to carry out feature point detection, to image I ₁, I ₂Detect respectively N ₁, N ₂Individual unique point; Adopt the Euclidean distance of proper vector to carry out the coupling of unique point as similarity measurement, obtain the feature point set of thick coupling; Adopt again RANSAC algorithm iteration purification match point; The model parameter of determining at last consists of perspective transition matrix H, and the unique point that consistent data is concentrated is to satisfying:

$\left(\begin{array}{c}{x}_i\\ y_i\\ 1\end{array}\right)=H*\left(\begin{array}{c}{x}_i^{\′}\\ y_i^{\′}\\ 1\end{array}\right)$

(x wherein _i, y _i) be I ₁In i (1≤i≤N ₁) individual unique point coordinate, (x ' _i, y ' _i) be image I ₂In characteristic of correspondence point coordinate with it, do not think that at the matching double points of most homogeneous data centralization the point of matching error is right;

(3) utilize the perspective transformation matrix that obtains in the Feature Points Matching process, piece image is mapped on the matched position of another width image;

$\left[\begin{array}{c}{x}^{\′}\\ y^{\′}\\ 1\end{array}\right]~H*\left[\begin{array}{c}x\\ y\\ 1\end{array}\right]$

What formula represented is with image I ₂Be mapped to image I ₁Matched position on, wherein [x ' y ' 1] is I ₂The homogeneous coordinates of middle pixel, [x y 1] is I ₂The homogeneous coordinates of the pixel after projective transformation;

(4) with image I ₁With image I ₂Image after projective transformation carries out the image difference computing;

(5) use self-adaption thresholding method, image after calculus of differences is carried out binary conversion treatment;

(6) there are a plurality of connected domains in binary image, for extracting its feature, carry out zone marker;

(7) at last by shape criterion locking " opal " effect target.

2. " opal " effect Target Recognition Algorithms of active imaging and passive imaging of common aperture as claimed in claim 1 system, it is characterized in that: in process of image registration, thick matching algorithm is improved, for the active imaging and passive imaging of common aperture system, dwindled from the ranks both direction hunting zone of seeking unique point to be matched the second width image.

Images