CN111340134B - A Fast Template Matching Method Based on Local Dynamic Warping - Google Patents

Abstract

The invention discloses a fast template matching method based on local dynamic regularization, which can be applied to the fields of workpiece positioning, industrial sorting, target tracking and the like. The steps are as follows: using the improved ring projection method (IRPT) to extract the feature vectors of the template image and the test subgraph, then initially estimate the similarity, screen out the candidate test subgraphs, and then use the proposed local dynamic warping method (LDTW) to calculate the candidate Test the similarity and zoom coefficient of the sub-images; take the test sub-image with the highest similarity value, and based on its corresponding zoom coefficient, cut out the smallest area containing the target object at the corresponding position of the test image, and finally use the direction code method (OC) Calculate the rotation angle of the area. Compared with the prior art, the present invention only needs one template image to calculate the scaling factor and the rotation angle, which solves the problem that the conventional algorithm needs a large number of template images with different combinations of the scaling factor and the rotation angle to calculate the scaling factor and the rotation angle. The algorithm is greatly simplified.

Description

A Fast Template Matching Method Based on Local Dynamic Warping

technical field

The invention relates to the technical field of machine vision positioning, in particular to a fast template matching method based on local dynamic regularization.

Background technique

Template matching algorithm is a key technology in machine vision. It uses a given template image to identify and locate similar test sub-images (ie target workpieces) in the test image. It is widely used in workpiece positioning systems and product quality inspection systems. Applications. As the working conditions become more and more complex, higher requirements are placed on the real-time and robustness of the template matching algorithm.

Conventional template matching methods usually face the challenges of scaling, rotation, noise, and illumination changes, and there is still a lack of good solutions. For example, Patent Document 1 (CN105046271A) discloses a method for positioning and detecting MELF components based on template matching. A large number of template images are obtained by rotating and scaling the original template image, and then each template image is used to match the test image one by one, so as to achieve the recognition of rotation The purpose of the angle and scaling factor, but the algorithm complexity is high; another example is Patent Document 2 (CN108805220A) discloses a fast template matching algorithm based on gradient integration, which combines image pyramids, extraction of contour points and gradient integration to reduce the number of The data calculation is redundant, but the algorithm still relies on a large number of template images with different combinations of scaling and rotation angles to identify the scale factor and rotation angle of the workpiece; another example is Patent Document 3 (CN102254181A) discloses a multi-order differential ring template matching Tracking method, this method realizes the calculation of the rotation angle of the target object based on the ring template matching criterion, but still cannot calculate the scale factor of the target object, and the method is not robust to illumination changes. To sum up, there is still a lack of a template matching method that can calculate the rotation angle and the scaling coefficient at the same time, and has low algorithm complexity and good robustness.

SUMMARY OF THE INVENTION

In order to improve the real-time performance and robustness of the template matching algorithm and simplify the algorithm at the same time, the present invention provides a fast template matching method based on local dynamic warping.

The technical scheme adopted in the present invention is as follows:

A fast template matching method based on local dynamic warping, the steps are as follows:

Step 1. Traverse the test graph, extract the test subgraph with the same size as the template image, and utilize the ring projection algorithm to extract the ring projection feature vector of the test subgraph and the template image;

Step 2. Take the ring projection feature vector obtained in step 1 as an input, calculate the rough estimation similarity of the test subgraph and the template image, and filter out the test subgraph whose similarity is greater than the No. 1 setting threshold, and list it as a candidate test subgraph;

Step 3. For the candidate test subgraph obtained in step 2, utilize the local dynamic regularization algorithm to calculate similarity and image scaling factor by the curve outline of the local alignment ring projection feature vector;

Step 4. After the test map traversal is completed, take the maximum value of the similarity. If the maximum value of the similarity is greater than or equal to the threshold set by No. 3, the coordinates of the corresponding test sub-map are the target position, and at the same time according to the corresponding zoom The coefficients crop out the smallest area containing the target object from the test map;

Step 5. Use the direction code algorithm to extract the minimum area obtained in step 4 and the direction code feature vector of the template image, then calculate the rotation angle of the image based on the direction code feature vector, and finally obtain the target position, scaling factor and rotation angle.

Preferably, the ring projection algorithm in the step 1 is as follows: the size of the template image is denoted as M×N, the center point (x ₀ , y ₀ ) of the template image is used as the origin to establish a polar coordinate system, and any pixel is denoted as T(r , θ), the ring projection eigenvector is denoted as IRPT,

R_max＝min(M/2，N/2)，s(r)是半径为r的圆环上的像素个数，T_min(r，θ)是圆环上所有像素强度的最小值。in,

R _max =min(M/2, N/2), s(r) is the number of pixels on a circle with radius r, and T _min (r, θ) is the minimum value of all pixel intensities on the circle.

Preferably, the algorithm for roughly estimating the similarity between the test subgraph and the template image in the step 2 is as follows: the ring projection feature vector extracted from the test submap is denoted as S, the ring projection feature vector extracted from the template image is denoted as T, and the test The rough estimation similarity between the sub-image and the template image is denoted as K _c , the larger the K _c is, the more similar the images are.

where n is the dimension of the vector X and S[0:m/2] is the first m/2 dimension of the feature vector S.

Preferably, the No. 1 set threshold in the step 2 is denoted as β ₁ , and takes 0.35≤β ₁ ≤0.5.

Preferably, the method 1 and the method 2 in the step 3 are to calculate the similarity and the image scaling coefficient by finding the optimal local matching relationship between the candidate test subgraph and the ring projection feature vector of the template image.

Preferably, a specific method in the step 3 is:

Step 3.1a. The ring projection feature vector extracted from the candidate test subgraph is denoted as S, and the ring projection feature vector extracted from the template image is denoted as T, input S and T, whose dimensions are m _s and m _t respectively, create a distance Matrix D and a cumulative distance matrix D ^acc , whose dimensions are both m _t ×m _s , initialize the distance function DIS=|xy|;

Step 3.2a. utilize the distance function DIS to calculate the distance between each element of the eigenvectors S and T, thereby obtaining the distance matrix D, then assign the distance matrix D to the cumulative distance matrix D ^acc ;

Step 3.3a. Use the following formula to update each element value of the cumulative distance matrix D ^acc , and obtain the cumulative distance matrix D ^acc after the update is completed,

Step 3.4a. For the last column of the cumulative distance matrix D ^acc , search for the element with the smallest value from bottom to top, its value is denoted as temp ₁ , and the position of this element is denoted as (i1, m _s );

Step 3.5a. For the last row of the cumulative distance matrix D ^acc , search for the element with the smallest value from right to left, its value is denoted as temp ₂ , and the position of this element is denoted as (i2, m _s );

Step 3.6a. The similarity between feature vectors S and T is recorded as K _s , and the scaling factor is recorded as K,

If temp ₁ is less than or equal to temp ₂ , then

If temp ₁ is greater than temp ₂ , then

Preferably, another specific method in the step 3 is:

Step 3.1b. Use Gaussian filtering to smooth and denoise the characteristic curve, the RPT eigenvector is denoted as f(x), the Gaussian function is denoted as g(x, σ), and the filtered RPT eigenvector F(x) is

Step 3.2b. The convolution kernel is denoted as T, then the discrete slope curve sequence F'(x) is

所对应的缩放系数k即为所求缩放系数K，Step 3.3b. The slope curve sequences obtained from the template image and the test sub-image are respectively denoted as T' and S', the scaling coefficient is denoted as k, the initial calculation range of the scaling coefficient is [k ₁ , k ₂ ], and the calculation accuracy of the scaling coefficient is ( Step size) is k', and the similarity K _s corresponding to each scaling coefficient k is calculated by the following formula, then the maximum similarity

The corresponding scaling factor k is the desired scaling factor K,

β₂是二号设定阈值，取10≤β₂≤15。where n _max =min(t, k×t),

β ₂ is the No. 2 set threshold, and takes 10≤β ₂ ≤15.

Preferably, the set threshold value No. 3 in the step 4 is denoted as β ₃ , and takes 0.55≤β ₃ ≤0.7.

Preferably, the direction code algorithm in the step 5 is a fan-shaped sampling method, that is, the image is divided into n fan-shaped areas, and then the intensity values of all pixels in the fan-shaped area are averaged and used as an element of the direction code feature vector, thus Get an orientation code eigenvector associated with the rotation angle.

Preferably, the step 5 is specifically:

Step 5.1. The size of the input image I is denoted as M×N, and the polar coordinate system is established with the center point (x ₀ , y ₀ ) of the input image as the origin, then any pixel can be expressed as I(r, θ), and the initialization direction code The angle calculation accuracy of the method is θ′, and the calculation formula of the direction code eigenvector OC is as follows:

r_max＝min(M/2，N/2)，s_r是落入扇形区域内的像素数量；in,

r _max =min(M/2, N/2), s _r is the number of pixels falling into the sector area;

计算公式如下：Step 5.2. The input image is a template image, denoted as T, and each θ corresponds to a direction code feature vector. Using the calculation method in step 5.1, the 360°/θ′ direction code feature vectors corresponding to the template image can be obtained, then the angle θ The corresponding eigenvectors are expressed as

Calculated as follows:

Wherein, n _max =360°/θ′-1;

计算公式如下所示：Step 5.3. The input image is the minimum area, denoted as S, and the calculation method in step 5.1 is used to obtain a direction code feature vector corresponding to the minimum area, which is expressed as

The calculation formula is as follows:

和特征向量

的相似度K_(θ，0°)，计算公式如下所示：Step 5.4. Calculate the eigenvectors corresponding to each θ

and eigenvectors

The similarity K _{(θ, 0°)} , the calculation formula is as follows:

β₄是四号设定阈值，取15≤β₄≤25；in,

β ₄ is the No. 4 set threshold, which is 15≤β ₄ ≤25;

Step 5.5. Find the maximum similarity K _{(θ, 0°)} , and the corresponding θ is the counterclockwise rotation angle of the minimum area relative to the template image. The calculation formula is as follows:

The beneficial effects of the present invention are:

(1) Conventional template matching methods require a large number of template images with different combinations of scaling and rotation angles to build a template library, and then matching a large number of template images and test images one by one to identify the scale factor and rotation angle of the workpiece. The matching process not only It is very time-consuming, and its calculation accuracy depends on the step size of the scaling factor and rotation angle. The present invention uses the proposed Local Dynamic Time Warping (LDTW, Local Dynamic Time Warping) algorithm to subtly solve this problem, and only needs a template image to calculate the zoom coefficient and the target position at the same time, and then uses the direction code (OC, Orientation Codes) algorithm method to calculate the rotation angle. Compared with the conventional template matching method, the present invention has very low algorithm complexity, and effectively improves the real-time performance of identification and positioning. In addition, the proposed LDTW method has good illumination robustness and has better stability.

(2) Conventional template matching methods usually suffer from noise and illumination changes, which in turn lead to unstable positioning accuracy, which is easy to cause production accidents in industrial sites. The present invention adopts the proposed improved ring projection (IRPT, Improved Ring Projection Transformation) algorithm to extract features, and the features have good noise robustness.

(3) The algorithm of the present invention can perform parallel calculation in the process of traversing the test graph to calculate the similarity between the test subgraph and the template image and using the OC feature vector to calculate the rotation angle, so as to meet the real-time requirements of industrial applications.

Description of drawings

FIG. 1 is a flowchart of a template matching method in an embodiment of the present invention.

FIG. 2 is a schematic diagram of a visual image processing process in an embodiment of the present invention (method 1).

FIG. 3 is a schematic diagram of a visual image processing process in an embodiment of the present invention (method 2).

FIG. 4 is a result diagram of processing various scaling coefficients and rotation angles by using a template image according to an embodiment of the present invention (method 1).

FIG. 5 is a result diagram of processing various scaling coefficients and rotation angles by using a template image according to an embodiment of the present invention (method 2).

Detailed ways

The present invention will be further described below with reference to the embodiments and the accompanying drawings. This embodiment provides a fast template matching method based on local dynamic warping, as shown in FIGS. 1 to 3 , and the steps are as follows.

Step S1. Traverse the test graph, extract the test subgraph with the same size as the template image, and use the ring projection algorithm to extract the ring projection feature vector of the test subgraph and the template image.

S1.1. Traverse the test image from top to bottom and from left to right, crop the test sub-image with the same size as the template image, and take the pixel coordinates of the center point of the test sub-image as the position coordinates of the test sub-image.

S1.2. Use the proposed improved Ring Projection (IRPT, Improved Ring Projection Transformation) algorithm to extract the IRPT feature vector of the test subgraph and the template image; x ₀ , y ₀ ) is the origin to establish a polar coordinate system, any pixel is represented as T(r, θ), and the ring projection feature vector is represented as IRPT,

R_max＝min(M/2，N/2)，s(r)是半径为r的圆环上的像素个数，T_min(r，θ)是圆环上所有像素强度的最小值。in,

R _max =min(M/2, N/2), s(r) is the number of pixels on a circle with radius r, and T _min (r, θ) is the minimum value of all pixel intensities on the circle.

Step S 2. Take the ring projection feature vector of S1 gained as input, calculate the rough estimation similarity of test subgraph and template image, and filter out the test subgraph whose similarity is greater than No. 1 setting threshold, and be listed as candidate test subgraph, The first set threshold is denoted as β ₁ .

The algorithm for rough estimation of similarity is as follows: the ring projection feature vector extracted from the test subgraph is denoted as S, the ring projection feature vector extracted from the template image is denoted as T, the rough estimation similarity between the test subgraph and the template image is denoted as K _c , The larger the K _c , the more similar the images are.

where n is the dimension of the vector X and S[0:m/2] is the first m/2 dimension of the feature vector S.

Usually, if β ₁ is too large, the correct candidate test subgraph will be filtered out by mistake, and if it is too small, the preliminary screening function will be lost, and take 0.35≤β ₁ ≤0.5. If K _c is greater than or equal to the threshold β ₁ , the test subgraph is listed as a candidate test subgraph; if it is less than, the similarity is set to 0. It is worth pointing out that only using the first m/2 dimensions of the feature vector S to calculate the similarity is beneficial to filter the background noise caused by image scaling.

Step S3. For the candidate test subgraph obtained by S 2 , based on the rotation invariance and global contour invariance of the IRPT eigenvector, utilize the local dynamic warping (LDTW, Local Dynamic Time Warping) algorithm to project the curve of the eigenvector through the local alignment ring. contour to calculate similarity and image scaling factor. This embodiment proposes two implementation methods, namely method 1 and method 2. The essence is to calculate the similarity and the image scaling coefficient by finding the optimal local matching relationship between the candidate test subgraph and the ring projection feature vector of the template image.

Method 1 is specifically:

S3.1a. The ring projection feature vector extracted from the candidate test subgraph is denoted as S, and the ring projection feature vector extracted from the template image is denoted as T, input S and T, whose dimensions are m _s and m _t respectively, create a distance Matrix D and a cumulative distance matrix D ^acc , whose dimensions are both m _t ×m _s , initialize the distance function DIS=|xy|.

S3.2a. Use the distance function DIS to calculate the distance between each element of the eigenvectors S and T to obtain the distance matrix D, and then assign the distance matrix D to the cumulative distance matrix D ^acc .

S3.3a. Use the following formula to update each element value of the cumulative distance matrix D ^acc , and obtain the cumulative distance matrix D ^acc after updating,

S3.4a. For the last column of the cumulative distance matrix D ^acc , search for the element with the smallest value from bottom to top, the value of which is denoted as temp ₁ , and the position of this element is denoted as (i1, m _s ).

S3.5a. For the last row of the cumulative distance matrix D ^acc , search for the element with the smallest value from right to left, whose value is denoted as temp ₂ , and the position of this element is denoted as (i2, m _s ).

S3.6a. The similarity between feature vectors S and T is recorded as K _s , and the scaling factor is recorded as K,

If temp ₁ is less than or equal to temp ₂ , then

If temp ₁ is greater than temp ₂ , then

Method 2 is specifically:

S3.1b. Use Gaussian filtering to smooth and denoise the characteristic curve. The RPT feature vector is denoted as f(x), and the Gaussian function is denoted as g(x, σ). The filtered RPT feature vector F(x) is:

S3.2b. The convolution kernel is denoted as T, then the discrete slope curve sequence F'(x) is:

所对应的缩放系数k即为所求缩放系数K，S3.3b. The slope curve sequences obtained from the template image and the test sub-image are denoted as T' and S' respectively, the scaling coefficient is denoted as k, the initial calculation range of the scaling coefficient is [k ₁ , k ₂ ], and the calculation accuracy of the scaling coefficient is ( Step size) is k', and the similarity K _s corresponding to each scaling coefficient k is calculated by the following formula, then the maximum similarity

The corresponding scaling factor k is the desired scaling factor K,

β₂是二号设定阈值，取10≤β₂≤15。where n _max =min(t, k×t),

β ₂ is the No. 2 set threshold, and takes 10≤β ₂ ≤15.

Step S 4. After the traversal of the test graph is completed, take the maximum value of the similarity. If the maximum value of the similarity is greater than or equal to the threshold set by No. 3, the coordinates of the corresponding test sub-image are the target position, and at the same time, according to the corresponding zoom factor, the smallest region (ROI, ROI, Region of interest); if it is less than, it means that there is no target workpiece in the test chart. The set threshold of No. 3 is recorded as β ₃ . If β ₃ is too large or too small, it will lead to false matching. Usually, 0.55≤β ₃ ≤0.7 is taken.

Step S 5. Utilize the direction code (OC) algorithm to extract the minimum area (ROI) obtained in S 4 and the direction code feature vector of the template image, then calculate the rotation angle of the image based on the direction code feature vector, and finally obtain the target position, scaling factor and Rotation angle.

The above direction code algorithm is a fan-shaped sampling method, that is, the image is divided into n fan-shaped areas, and then all pixel intensity values in the fan-shaped area are averaged and used as an element of the direction code feature vector, thereby obtaining a rotation angle. Direction code feature vector. Specifically:

S5.1. The size of the input image is recorded as M×N, and the polar coordinate system is established with the center point (x ₀ , y ₀ ) of the input image as the origin, then any pixel can be represented as T(r, θ); initialization direction code The angle calculation of the method is precise θ′, take θ′=1°, and the calculation formula of the direction code feature vector OC is as follows:

r_max＝min(M/2，N/2)，s_r是落入扇形区域内的像素数量。in,

r _max =min(M/2, N/2), s _r is the number of pixels falling within the sector.

计算公式如下：S5.2. The input image is a template image, denoted as T, and each θ corresponds to a direction code feature vector. Using the calculation method in step 5.1, the 360°/θ′ direction code feature vectors corresponding to the template image can be obtained, then the angle The eigenvector corresponding to θ is expressed as

Calculated as follows:

where n _max =360°/θ'-1.

计算公式如下所示：S5.3. The input image is the minimum area, denoted as S, and a direction code feature vector corresponding to the minimum area is obtained by using the calculation method in step 5.1, which is expressed as

The calculation formula is as follows:

和特征向量

的相似度K_(θ，0°)，计算公式如下所示：S5.4. Calculate the eigenvector corresponding to each θ

and eigenvectors

The similarity K _{(θ, 0°)} , the calculation formula is as follows:

β₄是四号设定阈值；β₄过大容易导致丧失计算旋转角度功能，过小则导致噪声鲁棒性不好，根据实际调参经验，β₄通常设置为15-25。in,

β ₄ is the No. 4 setting threshold; if β ₄ is too large, it will easily lead to the loss of the function of calculating the rotation angle, and if β 4 is too small, it will lead to poor noise robustness. According to the actual parameter adjustment experience, β ₄ is usually set to 15-25.

S5.5. Find the maximum similarity K _{(θ, 0°)} , and the corresponding θ is the counterclockwise rotation angle of the minimum area relative to the template image. The calculation formula is as follows:

Step S6. Use S5 to obtain the target position, zoom factor and rotation angle for workpiece positioning, industrial sorting or target tracking.

Obviously, the above-mentioned embodiments of the present invention are only examples for illustrating the present invention, rather than limiting the embodiments of the present invention. For those of ordinary skill in the art, changes or modifications in other different forms can also be made on the basis of the above description. All implementations need not and cannot be exhaustive here. And these obvious changes or changes derived from the essential spirit of the present invention still belong to the protection scope of the present invention.

Claims (5)

1. A fast template matching method based on local dynamic warping is characterized by comprising the following steps:

step 1, traversing the test graph, extracting a test subgraph with the same size as the template image, and extracting ring projection characteristic vectors of the test subgraph and the template image by using a ring projection algorithm;

the ring projection algorithm is as follows: the size of the template image is recorded as M multiplied by N, and the central point (x) of the template image is used ₀ ，y ₀ ) A polar coordinate system is established for the origin, any one pixel is denoted as T (r, theta), the ring projection feature vector is denoted as IRPT,

wherein,

R _max min (M/2, N/2), s (r) is the number of pixels on the circle with radius r, T _min (r, θ) is the minimum of all pixel intensities on the circle;

step 2, taking the ring projection feature vector obtained in the step 1 as input, calculating the roughly estimated similarity between the test subgraph and the template image, screening out the test subgraphs with the similarity larger than a first set threshold value, and listing the test subgraphs as candidate test subgraphs;

step 3, aiming at the candidate test subgraph obtained in the step 2, calculating the similarity and the image scaling coefficient by using a local dynamic warping algorithm and locally aligning the curve outline of the ring projection feature vector, namely, searching the optimal local matching relation of the candidate test subgraph and the ring projection feature vector of the template image and calculating the similarity and the image scaling coefficient;

a specific method of the step 3 is as follows:

step 3.1a, recording the ring projection characteristic vector extracted from the candidate test subgraph as S, recording the ring projection characteristic vector extracted from the template image as T, and inputting S and T, wherein the dimensionalities of the S and T are m respectively _s And m _t Creating a distance matrix D and a cumulative distance matrix D ^acc All dimensions of m _t ×m _s Initializing distance function DIS ═ x-y |;

step 3.2a, calculating the distance between each element of the eigenvector S and each element of the eigenvector T by using the distance function DIS to obtain a distance matrix D, and then assigning the distance matrix D to the accumulated distance matrix D ^acc ；

Step 3.3a. update the cumulative distance matrix D using the following formula ^acc After the updating of each element value, the accumulative distance matrix D is obtained ^acc ，

Step 3.4a. for the cumulative distance matrix D ^acc The last column of (1) searching the element with the smallest value from bottom to top, and recording the value as temp ₁ The position of the element is (i1, m) _s )；

Step 3.5a. for the cumulative distance matrix D ^acc The last line of (1) searches for the element with the smallest value from right to left, and the value is recorded as temp ₂ The position of the element is (i2, m) _s )；

Step 3.6a. the similarity of the characteristic vectors S and T is recorded as K _s The scaling factor, denoted as K,

if temp. is lower than the normal ₁ Temp less than or equal to ₂ Then, then

If temp. is lower than the normal ₁ Greater than temp ₂ Then, then

The other specific method of the step 3 is as follows:

step 3.1b, smoothing and denoising the characteristic curve by Gaussian filtering, wherein a ring projection characteristic vector is recorded as f (x), a Gaussian function is recorded as g (x, sigma), and a filtered ring projection characteristic vector F (x) is

Step 3.2b. the convolution kernel is marked as T, then the discrete slope curve sequence F' (x) is

And 3.3b, respectively recording the slope curve sequences obtained from the template image and the test subgraph as T 'and S', recording the scaling coefficient as k, and initializing the calculation range of the scaling coefficient as k ₁ ，k ₂ ]The calculation precision of the scaling coefficient is K', and the similarity K corresponding to each scaling coefficient K is calculated by using the following formula _s Then the maximum similarity

The corresponding scaling factor K is the desired scaling factor K,

wherein n is _max ＝min(t，k×t)，

β ₂ Setting a threshold value of 10-beta ₂ ≤15；

Step 4, after traversing the test chart, taking the maximum value of the similarity, if the maximum value of the similarity is greater than or equal to a third set threshold value, determining the coordinate of the corresponding test subgraph as the target position, and cutting out the minimum area containing the target object from the test chart according to the corresponding scaling coefficient;

step 5, extracting the direction code characteristic vectors of the minimum region and the template image obtained in the step 4 by using a direction code algorithm, and then calculating the rotation angle of the image based on the direction code characteristic vectors to finally obtain a target position, a scaling coefficient and the rotation angle; the direction code algorithm is a sector sampling method, namely, an image is divided into n sector areas, all pixel intensity values in the sector areas are averaged and used as an element of a direction code feature vector, and therefore a direction code feature vector related to a rotation angle is obtained.

2. The fast template matching method based on local dynamic warping as claimed in claim 1, wherein the algorithm for testing the rough estimated similarity between the sub-graph and the template image in step 2 is as follows: the ring projection characteristic vector extracted from the test subgraph is marked as S, the ring projection characteristic vector extracted from the template image is marked as T, and the rough estimation similarity between the test subgraph and the template image is marked as K _c ，K _c The larger the size the more similar the image is,

where n is the dimension of vector X, S [ 0: m/2 is the first m/2 dimension of the feature vector S.

3. The fast template matching method based on local dynamic warping as claimed in claim 1 or 2, wherein: setting a threshold value in the step 2 as beta ₁ Taking beta of not more than 0.35 ₁ ≤0.5。

4. The fast template matching method based on local dynamic warping as claimed in claim 1, wherein: setting a threshold value beta in the third step of the step 4 ₃ Taking beta not more than 0.55 ₃ ≤0.7。

5. The fast template matching method based on local dynamic warping as claimed in claim 1, wherein the step 5 specifically comprises:

step 5.1, recording the size of the input image I as M multiplied by N, and inputting the central point (x) of the image ₀ ，y ₀ ) Establishing a polar coordinate system for the origin, any one pixel can be represented as I (r, θ), the angle calculation precision θ' of the direction code method is initialized, and the direction code feature vector OC is calculated as follows:

wherein,

r _max ＝min(M/2，N/2)，s _r is the number of pixels falling within the sector area;

step 5.2, the input image is a template image and is marked as T, each theta corresponds to a direction code feature vector, 360 degrees/theta' direction code feature vectors corresponding to the template image can be obtained by utilizing the calculation method in the step 5.1, and the feature vector corresponding to the angle theta is expressed as

The calculation formula is as follows:

wherein n is _max ＝360°/θ′-1；

Step 5.3, the input image is a minimum area and is marked as S, and a direction code characteristic vector corresponding to the minimum area is obtained by using the calculation method in the step 5.1 and is expressed as

The calculation formula is as follows:

step 5.4, calculating the characteristic vector corresponding to each theta

And feature vectors

Similarity K of (2) _(θ，0°) The calculation formula is as follows:

wherein,

β ₄ setting a threshold value of 15-beta ₄ ≤25；

Step 5.5, find the biggest similarity K _(θ，0°) And the corresponding theta is the counterclockwise rotation angle of the minimum area relative to the template picture, and the calculation formula is as follows:

Images