CN103020967A - Unmanned aerial vehicle aerial image accurate matching method based on island edge characteristics - Google Patents

Abstract

The invention provides an unmanned aerial vehicle aerial image accurate matching method based on island edge characteristics and relates to the field of computer graphics. The unmanned aerial vehicle aerial image accurate matching method solves the problem of dislocation easily occurring during splicing of unmanned aerial vehicle aerial images for island edges. Anisotropic Gaussian filters and weighted median filters in weight dynamic allocation are adopted to filter noises in island aerial images, a chordal distance fixed point self-adaption linear fitting method is utilized to detect edge angular points, a maximum likelihood estimation sample consensus (MLESAC) algorithm is utilized to screen matching points, the screened matching points are substituted into an affine transformation model to solve parameters, and matched modified functions are utilized to perform accurate registration on the images. According to the unmanned aerial vehicle aerial image accurate matching method, island edge details are effective reserved, the extracted island edges are clear, and the edge angular points can be extracted rapidly. A screening and correction process of the angular points is designed, the island images are matched accurately, and the matching effect is good.

Description

Unmanned plane Aerial Images exact matching method based on the island edge feature

Technical field

The present invention relates to the field of computer graphics, refer more particularly to the extraction of picture edge characteristic, the identification of image border angle point and the accuracy registration between image.

Background technology

The edge is the separatrix of zones of different, is the set of significant pixel of image local Strength Changes.Rim detection is the basic problem of image processing and computer vision field as a requisite link in the feature extraction.Image Edge-Detection not only can greatly reduce the important feature information that data volume that image processes can keep again image.Rim detection can be divided into two types substantially: based on the rim detection of searching with based on zero rim detection of passing through.Mainly be to realize by the extreme value of finding the solution in the image first order derivative based on the rim detection of searching, usually with the direction place of boundary alignment in the gradient maximum; Realize based on the zero crossing that zero rim detection of passing through often adopts Laplacian zero crossing or nonlinear difference to represent.Edge detection algorithm commonly used has: Roberts operator, Prewitt operator, Sobel operator, LOG operator, Canny operator.

Angle point is the point of curvature value maximum on the boundary curve in the image, and violent variation can occur near the gray-scale value this point.Present Corner Detection Algorithm can be summarized as 3 classes: the Corner Detection of intensity-based image, based on the Corner Detection of bianry image, based on the Corner Detection of contour curve.The image-based Corner Detection is global search basically, and is less based on the Corner Detection data volume of edge contour, to the flase drop of angle point and undetected better than direct image-based method.

Because the Gaussian noise that the noise in the Aerial Images of island has the factors such as impulsive noise that the factors such as the sea is reflective form and sea cloud and mist to form.The filtering characteristic of conventional arithmetic operators can not meet the demands, and often causes the island edge extracting unintelligible, and match point is chosen the larger problem of error.And the correction to match point is perfect not in the subsequent process, causes the splicing of island edge pattern to tend to occur misplacing or missing the phenomenon of mating.At present image split-joint method is mainly for the land Aerial Images with obvious match point, substantially do not have for the image split-joint method at edge, island, and now increasing to the demand of island Image Mosaics.

Summary of the invention

The present invention can overcome the mistake matching problem that island edge image splicing occurs, a kind of unmanned plane Aerial Images exact matching method based on the island edge feature is provided, it has realized the complete clear extraction at edge, island, the fast detecting of edge angle point and the screening of follow-up angle point, match point correction flow process.

For achieving the above object, the present invention adopts following technical scheme, comprising following steps:

(1) with the weighted median filtering of weights dynamic assignment and impulsive noise and the Gaussian noise in the Aerial Images of anisotropic Gaussian filtering filtering island, carries out the island edge extracting;

(2) detect edge, island angle point based on " string is apart from the Adaptive Line Approximation of fixed point " method;

(3) with the MLESAC algorithm matching double points is screened;

(4) match point substitution affine Transform Model, find the solution parameter;

(5) set up the coupling correction function by the relaxative iteration method, image is carried out exact matching.

The present invention is based on the scrambling at edge, island, utilize the reflective impulsive noise that causes in improved weighted median filtering filtering sea, and the Gaussian noise that caused by cloud and mist of the gaussian filtering filtering that utilizes anisotropic, can effectively keep the island edge details, extract clearly edge, island.Utilize the method for " string is apart from the Adaptive Line Approximation of fixed point " that the edge, island is carried out fast angle point extraction, substitution affine Transform Model after through the MLESAC algorithm angle point being screened, and through the correction of coupling correction function, image is mated more accurately.

In the described step (1), utilize a 3*3 window that image pixel is added up, and utilize statistics that the weights of weighted median filtering and the yardstick of anisotropic gaussian filtering are carried out weight allocation, use finite difference computed image gradient and the direction of single order partial derivative after the filtering, and the non-maximum value of Grad suppressed, then dual threshold is complementary adjusts the final clear extraction that realizes the edge, island.

In the described step (2), detect edge, island angle point based on " string is apart from the Adaptive Line Approximation of fixed point " method, it comprises following performing step:

(a) determine the mobile increment factor s of string, any point P on the edge line _iConnect P _iAnd P _I+sForm a string C _i

(b) string C _iEach some P on the corresponding camber line _j, j ∈ [i, i+s) to string C _iDo vertical line, and calculate P _jTo string C _iVertical range d (P _j, C _i), to d (P _j, C _i) ask first order derivative, find the solution d ' (P _j, C _iAll P of)=0 _j, j ∈ [i, i+s) put as candidate angular;

(c) record string C _iEnd points, string with corresponding arc intersection point with (b) in try to achieve d ' (P _j, C _iAll P of)=0 _jPoint is as string C _iCandidate angular on institute's corresponding edge arc is to P _iCarry out the movement of increment factor s, form new string C _I-s=P _I-sP _iAnd C _I+s=P _I+sP _I+2s, repeat step in (b), until edge line stops;

(d) to (b) (c) each angle point in the candidate angular set that forms of step carry out fitting a straight line, determine whether it is angle point.

Adopt the MLESAC algorithm to filter out match point substitution affine Transform Model in described step (3), (4), (5), find the solution change of scale, the anglec of rotation, side-play amount, and utilize coupling correction function fine setting match point, carry out exact matching.

Beneficial effect of the present invention is: based on the unmanned plane Aerial Images exact matching method of island edge feature, impulsive noise in edge, the island Aerial Images and effective filtering of Gaussian noise have been realized, better kept edge details, edge carries out the angle point extraction quickly, and to the angle point screening of extracting, correct, solve efficiently the mistake matching problem of taking photo by plane and occur in the figure splicing in edge, unmanned plane island.

Description of drawings

Fig. 1 is two edges, island to be matched figure that take photo by plane;

Fig. 2 is for extracting the profile at edge, island;

Fig. 3 is that string is apart from the schematic diagram of the Adaptive Line Approximation of fixed point;

Fig. 4 is the extraction figure one of edge, island angle point;

Fig. 5 is the extraction figure two of edge, island angle point;

Fig. 6 is take photo by plane splicing effect after the picture coupling of island.

Embodiment

The present invention comprises following implementation step mainly for the splicing of edge, unmanned plane island Aerial Images:

(1) with the weighted median filtering of weights dynamic assignment and impulsive noise and the Gaussian noise in the Aerial Images of anisotropic Gaussian filtering filtering island, carries out the island edge extracting;

(2) detect edge, island angle point based on " string is apart from the Adaptive Line Approximation of fixed point " method;

(3) with the MLESAC algorithm matching double points is screened;

(4) match point substitution affine Transform Model, find the solution parameter;

(5) set up the coupling correction function by the relaxative iteration method, image is carried out exact matching.

1. island edge extracting, concrete steps are as follows:

(a) yardstick of weighted median filter weights, anisotropic Gaussian filtering is determined: the window with 3*3 is added up image pixel, and the pixel value of each grid is arranged { f from small to large ₀, f ₁, f ₂, f ₃, f ₄, f ₅, f ₆, f ₇f ₈; If f ₈-f ₀Difference greater than setting threshold, then the weight of intermediate value is strengthened, if then reduce the weight of intermediate value less than threshold value, make the weights real-time update of weighted median filtering.Yardstick on all directions of the anisotropic Gaussian filtering also pixel value of corresponding each grid is given corresponding scale-value;

(b) according to the weights of determining in (a) and scale-value the island Aerial Images is carried out the filtering processing;

(c) Grad of computed image each point and gradient direction, gradient magnitude

Reacted the edge strength of image, gradient direction α [x, y]=arctan (G _x(x, y)/G _y(x, y)) reacted the direction at edge, H[x, y] get the direction angle alpha [x, y] of local maximum, be exactly the edge direction of this point;

(d) window with 3*3 suppresses the non-maximum value of gradient, and carries out rim detection with dual threshold, finally obtains clearly edge, island.As shown in Figure 1 image is carried out edge extracting, and the result as shown in Figure 2.

2. edge, island Corner Detection, concrete steps are as follows:

(a) determine the mobile increment factor s of string, any point P on the edge line _iConnect P _iAnd P _I+sForm a string C _i, as shown in Figure 3;

(b) string C _iEach some P on the corresponding camber line _j, j ∈ [i, i+s) to string C _iDo vertical line, and calculate P _jTo string C _iVertical range d (P _j, C _i), to d (P _j, C _i) ask first order derivative, find the solution d ' (P _j, C _iAll P of)=0 _j, j ∈ [i, i+s) put as candidate angular;

(c) record string C _iEnd points, string with corresponding arc intersection point with (b) in try to achieve d ' (P _j, C _iAll P of)=0 _jPoint is as string C _iCandidate angular on institute's corresponding edge arc is to P _iCarry out the movement of increment factor s, form new string C _I-s=P _I-sP _iAnd C _I+s=P _I+sP _I+2s, repeat step in (b), until edge line stops;

(d) to (b) (c) each angle point in the candidate angular set that forms of step utilize curve trend before and after straight-line equation xsin β-ycos β=ρ match candidate angular, then least error distance $\mathrm{\ϵ}=\underset{i}{\mathrm{\Σ}}{(x_i\cos \mathrm{\β}+y_i\sin \mathrm{\β}-\mathrm{\ρ})}^2.$ Order $\frac{\∂\mathrm{\ϵ}}{\∂\mathrm{\ρ}}=0,$ $\frac{\∂\mathrm{\ϵ}}{\∂\mathrm{\β}}=0,$ Get $t_1=\underset{i}{\mathrm{\Σ}}({x_i}^2-{y_i}^2)-n({\stackrel{\‾}{x}}^2-{\stackrel{\‾}{y}}^2),$ $t_2=2(\stackrel{\‾}{\mathrm{xy}}-\underset{i}{\mathrm{\Σ}{x}_i{y}_i})$ (t ₁, t ₂Be not zero simultaneously), if | t ₁|＞| t ₂| then

Otherwise

Limit is according to the position (x of initial point ₀, y ₀) and the center of all discrete points By

Direction estimate the approximate range of β, and then it is adjusted in [0,2 π] scope.

(e) the β angle that obtains former and later two directions of candidate angular by (d) is respectively β ₁And β ₂, the declinate of former and later two directions is d _β=π-min{| β ₁-β ₂|, 2 π-| β ₁-β ₂|, declinate d _βBe directly proportional with curvature of a curve.Selected threshold Th _β, near the declinate the candidate point is judged, such as d _βLess than Th _β, then cast out this candidate point, final detected angle point such as Fig. 4, shown in Figure 5.

3. match point is screened

Maximum cycle K among the MLESAC is set, current cycle time j=1, and use the RANSAC algorithm of minimum intermediate value strategy to estimate the standard deviation of Gaussian distribution among the MLESAC; The initial matching point set is carried out stochastic sampling, the 3 pairs of same places of at every turn sampling, and calculate the accurate coefficient of an assembly, use the EM algorithm clearing mixing constant of this moment; Maximum iteration time T is set _Max=1024 and current cycle time j=1, and make initial mixing coefficient r ₀=0.5.Repeatedly initial point set is sampled, until cycle index j＞K, EM gets mixing constant r according to Bayes principle after finishing _I+1

4. find the solution affine Transform Model

According to image translation, rotation, scale transformation form, known affine Transform Model:

$\left[\begin{array}{c}x\\ y\end{array}\right]=k\left[\begin{array}{cc}\cos \mathrm{\θ}& \sin \mathrm{\θ}\\ -\sin \mathrm{\θ}& \cos \mathrm{\θ}\end{array}\right]\left[\begin{array}{c}{x}_0\\ y_0\end{array}\right]+\left[\begin{array}{c}\mathrm{\Δx}\\ \mathrm{\Δy}\end{array}\right],$

To two width of cloth Image estimation initial value k that will mate ₀, θ ₀, △ x ₀, Δ y ₀, set up the coordinate corresponding relation of two point sets, calculate two image boundary cross-correlation matrixs, find out the maximal correlation position, can obtain △ x ₀With Δ y ₀If k=is λ ₁k ₀, θ=λ ₁θ ₀, being principle to the maximum with cross-correlation coefficient and carrying out iterative search, self-adaptation obtains optimum value.In θ finds the solution, in 1 ' interval substitution model, according to maximal correlation principle iterative search, to obtain optimum value θ, to k ₀Disposal route similarly.After having obtained optimum value k and θ, again to △ x ₀With Δ y ₀Press the maximal correlation principle and search for by pixel along two orthogonal directionss of image, to obtain optimum value △ x ₀With Δ y ₀

5. set up the coupling correction function, carry out exact matching

Utilize the relaxative iteration method to set up coupling and correct function:

$M=\frac{1}{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{[f_1(x_i,y_i)-f_2(x_i+d,y_i)]}^2}+\frac{1}{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{[f_1(x_i,y_i)-f_2(x_i,y_i+d)]}^2}$

As a result figure in coupling randomly draws N to match point, calculates matching rate M, and M increases when adjusting x and y, at this moment should continue x and y to former direction adjustment; Otherwise, adjust round about; Constantly the iteration journey of correcting one's mistakes by the iteration of certain number of times, finally obtains the maximum value of M, and finish the adjustment of x and y this moment, obtains the take photo by plane exact matching of figure of edge, island, and splicing effect as shown in Figure 6.

Claims (6)

1. the unmanned plane Aerial Images exact matching method based on the island edge feature is characterized in that, may further comprise the steps:

(1) with the weighted median filtering of weights dynamic assignment and impulsive noise and the Gaussian noise in the Aerial Images of anisotropic Gaussian filtering filtering island, carries out the island edge extracting;

(2) detect edge, island angle point based on " string is apart from the Adaptive Line Approximation of fixed point " method;

(3) with the MLESAC algorithm matching double points is screened;

(4) match point substitution affine Transform Model, find the solution parameter;

(5) set up the coupling correction function by the relaxative iteration method, image is carried out exact matching.

2. the unmanned plane Aerial Images exact matching method based on the island edge feature according to claim 1, it is characterized in that, in the described step (1), the system of selection of weights dynamic assignment is as follows: according to the island local edge, window with 3*3 is added up image pixel, and the pixel value of each grid is arranged { f from small to large ₀, f ₁, f ₂, f ₃, f ₄, f ₅, f ₆, f ₇f ₈; If f ₈-f ₀Difference greater than setting threshold, then the weight of intermediate value is strengthened, if then reduce the weight of intermediate value less than threshold value; Yardstick on all directions of the anisotropic Gaussian filtering also pixel value of corresponding each grid is given corresponding scale-value.

3. the unmanned plane Aerial Images exact matching method based on the island edge feature according to claim 1 is characterized in that, in the described step (2), " string is apart from the Adaptive Line Approximation of fixed point " method step is as follows:

(a) determine the mobile increment factor s of string, any point P on the edge line _iConnect P _iAnd P _I+sForm a string C _i

(b) string C _iEach some P on the corresponding camber line _j, j ∈ [i, i+s) to string C _iDo vertical line, and calculate P _jTo string C _iVertical range d (P _j, C _i), to d (P _j, C _i) ask first order derivative, find the solution d ' (P _j, C _iAll P of)=0 o'clock _j, j ∈ [i, i+s) put as candidate angular;

(c) record string C _iEnd points, string with corresponding arc intersection point with (b) in try to achieve d ' (P _j, C _iAll P of)=0 _jPoint is as string C _iCandidate angular on institute's corresponding edge arc is to P _iCarry out the movement of increment factor s, form new string C _I-s=P _I-sP _iAnd C _I+s=P _I+sP _I+2s, repeat step in (b), until edge line stops;

(d) to (b) (c) each angle point in the candidate angular set that forms of step carry out fitting a straight line, determine whether it is angle point.

4. the unmanned plane Aerial Images exact matching method based on the island edge feature according to claim 1, it is characterized in that, in the described step (3), MLESAC screening match point method, estimate the standard deviation of Gaussian distribution among the MLESAC with the RANSAC algorithm of minimum intermediate value strategy, the initial matching point set is carried out stochastic sampling, 3 pairs of same places of each sampling, and calculate the accurate coefficient of an assembly, use the EM algorithm clearing mixing constant of this moment, repeatedly the initial matching point set is sampled, until exceed maximum cycle.

5. the unmanned plane Aerial Images exact matching method based on the island edge feature according to claim 1 is characterized in that, in the described step (4), finds the solution parameter according to affine Transform Model, treats that matching image is rotated, translation, scale transformation.

6. the unmanned plane Aerial Images exact matching method based on the island edge feature according to claim 1 is characterized in that, in the described step (5), sets up the coupling correction function by the relaxative iteration method, and fine setting match point coordinate carries out exact matching.

Images