CN101556695A - Image matching method - Google Patents

Abstract

The invention relates to an image matching method, comprising the following steps: (1) collecting images to be matched by an image collector and extracting template images from an image processor; (2) reducing the images to be matched and template images by half in the directions of width and height; (3) repeating step (2) until the images to be matched and template images reach specified sizes and carrying out gray scale normalization cross correlation matching on the reduced images to obtain the optimal matching positions (x, y); (4) returning to the previous images which are reduced by half and carrying out further matching on 9 search positions within the range of ((2x-1)-(2x+1), (2y-1)-(2y+1)) to obtain new optimal matching positions (x, y); (5) repeating step (2) until returning to the primary image, namely the original image, and carry out final matching to obtain the optimal matching positions and matching values. The method features shorter matching time and smaller amount of calculation.

Description

A kind of image matching method

Technical field

The invention belongs to graphic processing method, relate in particular to a kind of image matching method.

Background technology

Along with science and technology development, images match has become very important technology in the Image Information Processing field.Need to use image matching technology in about 40% the machine vision applications now, automatic monitoring, resource analysis, weather forecast, medical diagnosis, traffic administration, literal identification and the image retrieval etc. of the image tracking of terrain match, optics and the radar of related application from the industrial detection to the guided missile, the automatic monitoring of industrial flow-line, industrial instrument.Images match research has related to image acquisition, image pre-service, image segmentation, feature extraction etc., and combines closely with computer vision, multi-dimensional signal processing and numerical computation method etc.It also is some other image analysis technology, basis as stereoscopic vision, motion analysis, data fusion etc.: just because of the popularity of its application, new application and new requirement progressively produce, make the research of matching algorithm progressively move towards deeply the matching algorithm that occurred fast, stablize, robustness is good.Therefore, existing matching algorithm deployment analysis is had very important significance for actual engineering raising quality of image processing and accuracy of identification.

The normalized crosscorrelation matching algorithm is by the gray scale normalization cross-correlation coefficient of calculation template and image to be matched, and promptly cross correlation value is determined the degree of coupling.Cross-correlation method can overcome because tonal distortion that image acquisition and systematic error are brought and geometric distortion (translation distortion and rotation).

There is oversize shortcoming match time in traditional gray scale normalization cross-correlation coefficient matching algorithm.Its coupling and search position fixing process are all searching positions on the traversal detected image, calculate the correlation of each position, and the position of correlation maximum is a matched position.Be the product of two numerical value match time, and one is the number of pixels of hunting zone, and another is to calculate the time that the searching position related coefficient is spent, and the latter is depended on the size of template image, i.e. the number of pixels of template image.The shortcoming of this method is that the matching process calculated amount is big, needs the cost plenty of time, and the big more or detected image hunting zone of template image is big more, and the needed time is k ²l ²Multiple increase, k and l are respectively the increase multiple of template image and detected image.If all increasing, hunting zone and template image be that original twice, match time will be original 16 times.

Summary of the invention

Problem at prior art exists the invention provides the image matching method that a kind of match time is shorter, calculated amount is littler.

For achieving the above object, technical scheme of the present invention is: a kind of image matching method, and it may further comprise the steps: (1) utilizes image acquisition device to gather image to be matched, and accesses template image from image processor; (2) image to be matched and template image are all dwindled one times at width and short transverse; (3) repeating step (2) reaches the size of appointment up to matching image and template image, on the image that dwindles, use gray scale normalization simple crosscorrelation coupling obtain best match position (x, y); (4) turn back on the image that upper level dwindles a times (2x-1～2x+1,9 searching positions in the scope of 2y-1～2y+1) further mate again, obtain new best match position (x, y); (5) repeating step (4), up to turning back to first order image, promptly original image carries out last coupling, the best match position of acquisition and matching value.

In the above-mentioned steps (3), gray scale normalization simple crosscorrelation coupling is the gray matrix with a certain size each pixel of image to be matched, with template image might window the gray scale array, search for comparison by default method for measuring similarity, and carry out the cross correlation value coupling.

The computing formula of the place normalization cross-correlation coefficient of search window determining positions template image in image to be matched when cross correlation value is maximum is shown below,

$R(i,j)=\frac{\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}(C(x+i,y+j)-\stackrel{\‾}{C}(i,j))(P(x,y)-\stackrel{\‾}{P})}{\sqrt{\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}{(C(x+i,y+j)-\stackrel{\‾}{C}(i,j))}^2\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}{(P(x,y)-\stackrel{\‾}{P})}^2}}$

(i j) is illustrated on the detected image that (i is the image in the search window of initial point and the normalized crosscorrelation coefficient value of template image j), and image is meant the image block that has identical size on image to be detected with template image in the above-mentioned search window with point to R; (x+i is that coordinate is (x+i, y+j) gray values of pixel points on the detected image y+j) to C; (i is that (i j) is the average gray of image of the search window of initial point to some j) to C; (x is that coordinate is (x, y) gray values of pixel points on the template image y) to P; P is the average gray value of template image; The value of x and y is the coordinate figure of all pixels of template image; The span of i and j is the coordinate figure of all pixels of hunting zone.

The present invention is with respect to the advantage of prior art:

The match time of tradition gray scale normalization simple crosscorrelation matching algorithm is very big, the calculating that is cross correlation value on the one hand need use the gray-scale value of template image and all pixels of search window image to calculate, and is on the other hand to want interior all pixels of traversal search scope to calculate the cross correlation value of each searching position.This makes along with the increase of image, sharp increase match time.It is less that the present invention dwindles the back image, and the coupling calculated amount also reduces rapidly, can effectively shorten match time.

Description of drawings

Fig. 1 is stacked two minutes simple crosscorrelation matching algorithm synoptic diagram of the present invention.

Embodiment

The invention provides a kind of image matching method, as shown in Figure 1, it may further comprise the steps:

(1) utilizes image acquisition device to gather image to be matched, and from image processor, access template image;

(2) image to be matched and template image are all dwindled one times at width and short transverse;

(3) repeating step (2) reaches the size of appointment up to matching image and template image, on the image that dwindles, use gray scale normalization simple crosscorrelation coupling obtain best match position (x, y);

(4) turn back on the image that upper level dwindles a times (2x-1～2x+1,9 searching positions in the scope of 2y-1～2y+1) further mate again, obtain new best match position (x, y);

(5) repeating step (4), up to turning back to first order image, promptly original image carries out last coupling, the best match position of acquisition and matching value.

In the above-mentioned steps (3), gray scale normalization simple crosscorrelation coupling is the gray matrix with a certain size each pixel of image to be matched, with template image might window the gray scale array, search for comparison by default method for measuring similarity, and carry out the cross correlation value coupling.

The computing formula of the place normalization cross-correlation coefficient of search window determining positions template image in image to be matched when cross correlation value is maximum is shown below,

$R(i,j)=\frac{\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}(C(x+i,y+j)-\stackrel{\‾}{C}(i,j))(P(x,y)-\stackrel{\‾}{P})}{\sqrt{\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}{(C(x+i,y+j)-\stackrel{\‾}{C}(i,j))}^2\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}{(P(x,y)-\stackrel{\‾}{P})}^2}}$

(i j) is illustrated on the detected image that (i is the image in the search window of initial point and the normalized crosscorrelation coefficient value of template image j), and image is meant the image block that has identical size on image to be detected with template image in the above-mentioned search window with point to R; (x+i is that coordinate is (x+i, y+j) gray values of pixel points on the detected image y+j) to C; (i is that (i j) is the average gray of image of the search window of initial point to some j) to C; (x is that coordinate is (x, y) gray values of pixel points on the template image y) to P; P is the average gray value of template image; The value of x and y is the coordinate figure of all pixels of template image; The span of i and j is the coordinate figure of all pixels of hunting zone.

The present invention is with stacked two minutes simple crosscorrelation couplings of this method called after, and the meaning of half is repeatedly dwindled in expression.The advantage of stacked two minutes matching algorithms is to have reduced the hunting zone simultaneously and calculated the number of pixels that cross-correlation coefficient relates to, and therefore can Mrs dwindle match time.Suppose the time that the image when prime uses traditional simple crosscorrelation matching algorithm to need is Tt, the number of pixels of whole hunting zone is S, then the image of next stage time of using traditional simple crosscorrelation matching algorithm to need is T/16, then to use stacked two minutes matching methods to carry out two a fens required time when the prime image then is:

Tr(1)＝9Tt/S+Tt/16 (2)

Carrying out n two fens required time is:

Tr(n)＝9Tt/S+9*(Tt/16)/(S/4)+……+

9*(Tt/24(n-1))/(S/22(n-1))+Tt/24n (3)

Use one time two minutes stacked two minutes algorithms and use difference match time of traditional simple crosscorrelation matching algorithm to be working as prime as can be known by (2) formula:

ΔT＝(15/16-9/S)*Tt (4)

S is the number of pixels of the hunting zone when using traditional simple crosscorrelation matching algorithm when the prime image.If two fens number of times increase, the S value of the big more image of progression is more little, and Δ T is more near 0, when 15/16 ≈ 9/S, and promptly during S ≈ 9, Δ T ≈ 0.At this moment reach minimum value match time, promptly in theory, number of times was big more in two minutes, and the match time of algorithm is more little, when S ≈ 9, reaches minimum value.Though two minutes each time, image dwindled half, image dwindles can lose image information,, if comprise the image that will search in the image to be matched, because image and detected image to be matched is to dwindle simultaneously, still there is matching relationship in both.But in order to ensure the accuracy of matching result, the amount of decrease of adding the time when number of times was bigger in two minutes is also more and more littler, therefore can specify the mode of minimum image size to control two fens number of times.

Claims (3)

1, a kind of image matching method is characterized in that may further comprise the steps:

(1) utilizes image acquisition device to gather image to be matched, and from image processor, access template image;

(2) image to be matched and template image are all dwindled one times at width and short transverse;

(3) repeating step (2) reaches the size of appointment up to matching image and template image, on the image that dwindles, use gray scale normalization simple crosscorrelation coupling obtain best match position (x, y);

(4) turn back on the image that upper level dwindles a times (2x-1～2x+1,9 searching positions in the scope of 2y-1～2y+1) further mate again, obtain new best match position (x, y);

(5) repeating step (4), up to turning back to first order image, promptly original image carries out last coupling, the best match position of acquisition and matching value.

2, image matching method according to claim 1, it is characterized in that: in the above-mentioned steps (3), gray scale normalization simple crosscorrelation coupling is the gray matrix with a certain size each pixel of image to be matched, with template image might window the gray scale array, search for comparison by default method for measuring similarity, and carry out the cross correlation value coupling.

3, image matching method according to claim 2 is characterized in that: the computing formula of the place normalization cross-correlation coefficient of search window determining positions template image in image to be matched when cross correlation value is maximum is shown below,

$R(i,j)=\frac{\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}(C(x+i,y+j)-\stackrel{\‾}{C}(i,j))(P(x,y)-\stackrel{\‾}{P})}{\sqrt{\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}{(C(x+i,y+j)-\stackrel{\‾}{C}(i,j))}^2\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}{(P(x,y)-\stackrel{\‾}{P})}^2}}$

(i j) is illustrated on the detected image that (i is the image in the search window of initial point and the normalized crosscorrelation coefficient value of template image j), and image is meant the image block that has identical size on image to be detected with template image in the above-mentioned search window with point to R; (x+i is that coordinate is (x+i, y+j) gray values of pixel points on the detected image y+j) to C; (i is that (i j) is the average gray of image of the search window of initial point to some j) to C; (x is that coordinate is (x, y) gray values of pixel points on the template image y) to P; P is the average gray value of template image; The value of x and y is the coordinate figure of all pixels of template image; The span of i and j is the coordinate figure of all pixels of hunting zone.

Images