CN101226636B - An Image Matching Method of Rigid Body Transformation - Google Patents

Abstract

An image matching method applied to a rigid body transformation relational model. Including steps: regularly select a plurality of small template blocks (or feature points) in the current image, and judge the quality, and eliminate inappropriate ones (or weighted values according to the quality); respectively find the most matching corresponding ones in the previous image Position; according to the known geometric constraints, assign a weight to each coordinate point pair of the matching result; then fit the rigid body transformation relationship of the current image relative to the previous image. The method of the invention can effectively improve the robustness of the matching result under the condition of only adding a small amount of calculation. This method is applied to image matching in real-time ultrasound wide-field imaging to obtain high-quality imaging results.

Description

A kind of matching process of image of rigid body transformation relation

Technical field the present invention relates to image processing techniques, specifically, relates to the matching technique between the image that meets rigid body transformation relation.

Background technology is in the application of some images match, and the transformation relation between adjacent image meets rigid body transformation relation.In the ultrasonic scanning diagnosis, to move relative to even velocity, the displacement of adjacent two interframe is little, and is out of shape less in the target location for probe.If will generate wide field-of-view image by the image sequence that generates in the scanning process, consider the error accumulation effect in the coupling splicing, generally in coupling, use the rigid body translation model, think that promptly adjacent two interframe have only rotation and translation relation.In this class is used,, can obtain better corresponding between then adjacent two images if two adjacent images are used affined transformation model or multinomial model etc.But in the process that multiframe is constantly mated backward, the accumulation of distortion and error can make the image fault of back more and more serious, needs with the method for using global optimization.And in real-time occasion of mating joining image-forming, the method for global optimization can't be used smoothly.One of solution is to use the rigid body translation model in the coupling of adjacent two images.

Here the rigid body translation model of said image promptly has only the model of rotation and translation relation, and promptly the transformation relation of the relative previous image of present image is expressed as:

$\left(\begin{array}{c}X\\ Y\end{array}\right)=\left(\begin{array}{cc}\cos \mathrm{\θ}& -\sin \mathrm{\θ}\\ \sin \mathrm{\θ}& \cos \mathrm{\θ}\end{array}\right)\left(\begin{array}{c}{X}^{\′}\\ Y^{\′}\end{array}\right)+\left(\begin{array}{c}{X^{\′}}_0\\ {Y^{\′}}_0\end{array}\right)$

The coordinate on (X ', Y ') expression present image wherein, (X Y) is illustrated in the coordinate of previous image correspondence position, and θ represents the anglec of rotation, (X ' ₀, Y ' ₀) the expression translational movement.This transformation model has an important characteristic, at the relative position relation between point on the present image, can not change after conversion, remains unchanged as distance between two points etc.

Image matching method comprises that mainly two kinds of thinkings respectively have superiority based on the coupling of template with based on the coupling of feature.The purpose of these two kinds of thinkings all is to obtain a plurality of " point " coordinate corresponding relation between two adjacent images.According to specific model, utilize these " point " coordinate corresponding relations then, calculate the transformation relation of two integral image.When calculating whole corresponding relation, often use the least square method match by the corresponding relation of a plurality of sample points.A shortcoming of least square method is to depart from wholely when big as the individual samples point, and whole fitting result " is drawn " in meeting inclined to one side, and in the images match process, the situation that minority template (or unique point) mismatches appears in regular meeting.Fig. 1 has represented that wherein 1 be expected result with least square method match straight line, and 2 is fitting result, and to depart from integral body far away owing to a point, and make the result produce bigger deviation.Addressing this problem method commonly used is to iterate calculating, as M estimation and stochastic sampling coherence method etc.On the one hand, these method operands are bigger, and having limited it needs the application in the field of calculating in real time at some.On the other hand, they are universal methods, can use under various models, do not consider to utilize the characteristics of rigid body translation model to improve robustness.

Summary of the invention the present invention takes into full account under the rigid body translation model characteristics of corresponding relation between adjacent two images, the character of utilizing the relative position relation of point coordinate on the image under rigid body translation, to remain unchanged, significantly do not increasing under the condition of calculated amount, improving the robustness of matching result.The present invention has lower realization cost and lower computation complexity, makes method of the present invention can be used for the field of the less operand of needs, higher robustness, as the real-time wild eyeshot imaging of ultrasonic device.

Basic ideas of the present invention are: clocklike choose satisfactory template (or unique point) in present image, find corresponding coupling respectively in previous image, obtain the position corresponding relation; Utilize the geometrical constraint between the template of choosing (or unique point) then, judge the situation that departs from of corresponding position in the matching result; Again according to the degree that departs from for " point " of each correspondence to distributing different weights, go out the transformation relation of present image by these position corresponding relation weighted fittings to previous image.

The technical scheme that realizes thinking of the present invention is, utilizes the characteristics of rigid body translation, and a kind of image matching method of robust is provided, and only increases low computational effort relatively.Concrete steps are as follows:

A. a plurality of little template pieces (or unique point) are clocklike chosen in the fixed position in present image, and wherein second-rate (or by the quality tax weights) of rejecting;

B. each effective template piece (or unique point) finds the correspondence position that mates most in previous image;

C. according to known geometrical constraint, compose weights for the matching result of each effective template piece (or unique point);

D. by each correspondence position relation and weights, the transformation relation of the relative previous image of match present image;

It is characterized in that: the template piece (or unique point) in the steps A is in the fixed position, clocklike chooses.Geometrical constraint refers in particular among the step C: in the matching result, in the image with delegation or the same distance that lists other template matches position and current template matches position, with the deviation that distance should be arranged.Promptly concern that with position given in the steps A it is right to make matching result depart from actual positional relationship coupling far away " point ", accounts for less proportion when match as geometrical constraint, be used to the robustness that guarantees that the image rigid body transformation relation calculates.Adopt such scheme, can only increase relatively under the condition of low computational effort, effectively improve the robustness of matching result.

Deviation synoptic diagram when description of drawings Fig. 1 is the least square method fitting a straight line

A kind of template when Fig. 2 is to use template matches is chosen the mode synoptic diagram

Fig. 3 is the schematic flow sheet of a specific embodiment

Embodiment is described one embodiment of the present of invention now in conjunction with the accompanying drawings.

Here use the template matches mode as an object lesson, elaborate embodiments of the present invention.

When two images are mated, use the process of template matches to be exactly: at first on present image, to choose a plurality of little template pieces, search matched on previous image respectively, find best match position, go out transformation relation between two images with the coordinate Calculation of the coordinate of these little template pieces and corresponding matched position thereof then.The flow process of one embodiment of the present of invention as shown in Figure 3, concrete steps are as follows:

A. a plurality of little template pieces are clocklike chosen in zone to be matched in present image, and each template piece are carried out quality judge, reject second-rate template piece (or by quality tax weights);

B. each effective template piece search matched in previous image finds the correspondence position that mates most;

C. according to known geometrical constraint, compose weights for the matching result of each effective template;

D. the correspondence position by each effective template piece concerns and weights the transformation relation of the relative previous image of match present image;

Choosing as shown in Figure 2 the matching template piece in the steps A of present embodiment, wherein 3 represent present images, the a plurality of little image block of choosing on the 4 expression present images, position and the quantity chosen determine by actual conditions, can (be not limited to) be taken as the equidistant several row of several row in zone to be matched, so that calculate and obtain better stability.These little image blocks generally are taken as rectangle, also can get other shape.In the present embodiment, the wide height of each template piece can be taken as (pixel count) 16 * 16,32 * 32,48 * 48,64 * 64 etc.

The template piece of choosing will carry out quality to be judged, rejects second-rate piece (as too evenly, all pixels are near black or white etc.), to reduce erroneous matching in advance, strengthen and mate robustness.If the less weights of ropy tax, the big weights of the measured tax of matter, then the weights that use during overall fit are products of the matching result weights among quality weights and the step C.

Get each fixed effective template piece, respectively near the search matched correspondence position in previous image.In the present embodiment, use to calculate when the absolute difference of the pixel value of front template and previous image corresponding position and method, absolute difference and being expressed as

F wherein _iExpression is as the gray-scale value of i pixel of front template, g _iThe grey scale pixel value of corresponding position in the expression previous image.In whole hunting zone, the position of mating most when front template is thought in the position of sad value minimum in previous image.Remove the method in the present embodiment, also have (being not limited to) to use the method for related coefficient coupling, can obtain essentially identical result.

Each effective template piece calculates in previous image, finds one to think the position of coupling.Each effective template in present image coordinate and previous image in the coordinate of corresponding matched position to have constituted corresponding point right.Simulate the transformation relation of present image by the right relation of these points to previous image.Relation meets the rigid body translation model between two images, thus each point to coordinate transform and integral image transformation relation similar, be expressed as:

$\left(\begin{array}{c}{X}_i\\ Y_i\end{array}\right)=\left(\begin{array}{cc}\cos \mathrm{\θ}& -\sin \mathrm{\θ}\\ \sin \mathrm{\θ}& \cos \mathrm{\θ}\end{array}\right)\left(\begin{array}{c}{X^{\′}}_i\\ {Y^{\′}}_i\end{array}\right)+\left(\begin{array}{c}{X^{\′}}_0\\ {Y^{\′}}_0\end{array}\right)$

Wherein (X ' _i, Y ' _i) the effective coordinate of template in the expression present image, (X _i, Y _i) being illustrated in the previous image coordinate of corresponding matched position, θ represents the anglec of rotation, (X ' ₀, Y ' ₀) the expression translational movement.Final goal is exactly according to known a plurality of to coordinate, obtains the anglec of rotation and translational movement.

In the matching result of each effective template, have a few errors matching result sometimes, the coupling of these mistakes can reduce the robustness of fitting result.The present invention's known geometric relationship when choosing effective template composes weights for the matching result of each effective template, and match point is to participating in match by the weight size respectively.In the present embodiment, the distance between each effective template is known.Suppose to have in the present image position of two effective templates to be respectively a, b, their corresponding correct matched positions in previous image are A, B.Because use the rigid body translation model between two images, so the distance between A, B should equal the distance between a, b.If it is far away more that distance between A, B and due distance depart from, then the possibility of matching error is big more, and the weight in match is just more little.In the present embodiment, (being not limited to) defines the right weight of each point is the stack of Gaussian function:

$W_I=\underset{J}{\mathrm{\Σ}}{e}^{-{(\mathrm{diff}\_\mathrm{dist}(I,J))}^2/{\mathrm{\σ}}^2}$

Wherein the some J in the summation represent with I with delegation or at the point of same row, diff_dist (I, J) distance between expression I and J and their corresponding effective template i and the distance between j poor, σ has determined the A/F of Gaussian function, is used to regulate the difference degree of weight.Weight calculation result does not need high precision, as long as can distinguish the importance of each little template block search matching result, and therefore can be by discretize and calculate the method raising arithmetic speed of look-up table in advance.

Transformation relation between " point to " that use provides above and the right weight of each match point use weighted least-squares to simulate the rigid body transformation relation of present image and previous image.

Embodiments of the invention are realized on the PC platform, through experimental verification, under the condition that only increases low computational effort, can effectively improve the robustness of matching result.Simultaneously, this method is applied to the images match in the ultrasonic real-time wild eyeshot imaging, obtains high-quality imaging results.

Claims (2)

1. An image matching method applied to a rigid body transformation relation model, comprising steps: A. Select several rows and columns of template blocks equidistant in the area to be matched in the current image; after the template block is selected, perform quality judgment, and remove template blocks that are too uniform, all pixels are close to black, or all pixels are close to white; B. Each valid template block finds its best matching corresponding position in the previous image; C. Assign a weight to the matching result of each effective template block according to the known geometric constraints; D. by each corresponding position relation and weight, fitting current image relative to the transformation relation of previous image; It is characterized in that: The valid template block in steps B and C refers to the template block retained after the quality judgment in step A; In step C, the positional relationship given in step A is used as a geometric constraint, so that the matching result deviates from the "point" pair that is far from the actual positional relationship, which accounts for a small proportion in the fitting, and is used to ensure the robustness of the calculation of the image rigid body transformation relationship. Among them, the geometric constraint specifically refers to: for two template blocks on the same row or column in the current frame image, the correct matching positions corresponding to the two template blocks are calculated in the previous frame, and the correct matching position between The deviation between the distance of and the distance between the two template blocks in the current frame is a geometric constraint. 2. according to the image matching method described in claim 1, it is characterized in that: After the template blocks are selected in step A, the quality is judged, and a weight is assigned to each template block according to the quality; the weight used in step D is the product of the quality weight and the matching result weight in step C.

Images