CN116704220A - A Contour Target Recognition Method Based on the Combination of Centroid Height Increment and DTW - Google Patents

Abstract

The invention provides a contour target recognition method based on the combination of centroid height increment and DTW, and relates to the technical field of contour target recognition. The contour target recognition method based on the combination of centroid height increment and DTW includes the following steps: Find the centroid height increment feature matrix of the template image and the target image; use the DTW algorithm to calculate the centroid height increment feature matrix of the template image and the target image Compare to obtain the first feature distance; after flipping the target image, obtain the centroid height incremental feature matrix of the flipped image; compare the template image and the flipped image's centroid height incremental feature matrix through the DTW algorithm, and obtain the second Feature distance; use the minimum value function to obtain the minimum feature distance between the first feature distance and the second feature distance; introduce shape complexity into the minimum feature distance to obtain the similarity distance, and output the recognition result according to the similarity distance. Therefore, it can be ensured that the real-time performance of target recognition can be improved while the recognition rate effect is good.

Description

A Contour Target Recognition Method Based on the Combination of Centroid Height Increment and DTW

technical field

The invention relates to the technical field of target recognition, in particular to a contour target recognition method based on the combination of centroid height increment and DTW.

Background technique

In the field of object recognition, shape is an effective and highly utilized feature in object matching. Generally, shape representation methods include contour-based shape representation, region-based shape representation and skeleton-based shape representation. Contours are more and more used in object recognition methods because they have richer information, are not easily affected by changes in illumination, object color, and texture, and can effectively describe large-scale object structures.

In related technologies, some measure the similarity between contours through dynamic programming (Dynamic Programming, DP); some extract the skeleton features of the target contour, and then obtain the similarity between contours through the shape context algorithm; A shape similarity measurement method based on principal curvature enhanced distance transform for image recognition. However, there are technical problems such as high time complexity and low recognition processing speed in the above-mentioned schemes, which make the above-mentioned schemes not suitable for real-time target recognition.

Contents of the invention

The purpose of the embodiment of the present invention is to provide a contour target recognition method based on the combination of centroid height increment and DTW, so as to solve the technical problems that the contour target recognition method in the related art has low recognition processing speed and is not suitable for real-time target recognition .

In order to solve the above technical problems, embodiments of the present invention provide the following technical solutions:

The invention provides a contour target recognition method based on the combination of the centroid height increment and DTW, comprising the following steps: obtaining the centroid height increment feature matrix of the template image and the target image; Compare the centroid height incremental feature matrix to obtain the first feature distance; after flipping the target image, obtain the centroid height incremental feature matrix of the flipped image; use the DTW algorithm to compare the centroid height incremental feature matrix of the template image and the flipped image Perform comparison to obtain the second feature distance; use the minimum value function to find the minimum feature distance between the first feature distance and the second feature distance; introduce shape complexity into the minimum feature distance to obtain the similarity distance, according to the similarity distance Output the recognition result.

Further, obtaining the centroid height incremental feature matrix includes the following steps: preprocessing the image; extracting the peripheral contour of the image; extracting a plurality of sampling points on the peripheral contour to generate a contour sampling point set; calculating the contour of the contour sampling point set Centroid; calculate centroid height; calculate centroid height increment; obtain centroid height increment sequence; obtain centroid height increment matrix.

Further, preprocessing the image specifically includes the following steps: converting the image from a three-channel image to a single-channel image; performing thresholding noise reduction processing on the single-channel image; extracting the peripheral contour of the image includes the following steps: using Canny differential operator Extract the edge of the image; use the image morphology operation to expand the thin edge and fill the hole to form a complete peripheral contour; again use the Canny differential operator to accurately extract the peripheral contour.

Further, extracting a plurality of sampling points on the peripheral contour, and generating a contour sampling point set specifically includes: extracting N sampling points at equal intervals on the peripheral contour, and generating a contour sampling point set M; wherein, M={m _j }={ m ₁ , m ₂ ,..., m _N }; calculating the contour centroid of the contour sampling point set specifically includes: using the following formula to calculate the contour centroid Q(x ₀ , y ₀ ) of the contour sampling point set:

Calculating the centroid height specifically includes: defining the Euclidean distance between any sampling point m _j (x _j , y _j ) and the contour centroid Q(x ₀ , y ₀ ) as the centroid height of any sampling point m _j , j=( 1, 2,...N); use the following formula to calculate the centroid height g _j between any sampling point m _j (x _j , y _j ) and the contour centroid Q(x ₀ , y ₀ ): The calculation of the centroid height increment specifically includes: defining the difference between the centroid height g _j of any sampling point m _j and the centroid height gi of a certain sampling point m _i , which is any sampling point m _j relative to a certain sampling point The centroid height increment hi of m _{i , j} , i=(1, 2, ... N); obtaining the centroid height increment sequence specifically includes: calculating the centroid height of all sampling points relative to a certain sampling point _mi Increment, and arrange the centroid height increments of all sampling points relative to a certain sampling point m _i in the order of sampling points, and the centroid height sequence H _i of sampling point mi can be obtained: _{H i =} (h _{i, i} , h _{i, i+1} , h _{i, N} , h _{i, 1} ,...h _{i, i+1} ) ^T ; this sequence includes N elements, corresponding to the centroid height of N sampling points relative to sampling point m _i Increment; obtaining the centroid height increment matrix specifically includes: arranging the centroid height increment sequence H _i corresponding to all sampling points according to the order of the sampling points to obtain the centroid height increment matrix L (M): L (M)=(H ₁ , H ₂ , . . . , H _N-1 , H _N ).

Further, obtaining the centroid height increment feature matrix also includes the following steps: performing normalization processing on the centroid height increment h _{i, j} , to obtain the normalized centroid height increment ||h _{i, j} || is the modulus of the centroid height increment; the normalized centroid height increment sequence Add a positive integer coefficient d to the centroid height increment sequence _Hi after normalization processing, d (1<d<N); divide the centroid height increment sequence Hi after normalization processing into S disjoint sub- Sequence [1, d], [1+d, 2d]..., wherein, S=[N/d]; use the following formula to calculate the average value c _i,t of the centroid height increment of each subsequence:

Among them, t=1, 2,...S; arrange the S average value data in an orderly manner, and obtain the feature sequence C _i of the sampling point m _i after smoothing processing: C _i =(c _{i, 1} , c _{i , 2} ,..., ci _{, S-1} , ci _{, s} ) ^T ; the centroid height incremental feature sequence C _i after smoothing all sampling points is arranged according to the order of sampling points, and the smoothed centroid height is obtained Incremental feature matrix F(M): F(M)=(C ₁ , C ₂ , . . . , C _N-1 , C _N ).

Further, by using the DTW algorithm to compare the centroid height incremental feature matrix of the template image and the target image, obtaining the first feature distance specifically includes: setting the sample point set of the target image contour as M, M={m ₁ , m ₂ , ..., m _N }; set the template image contour sampling point set as Z, Z={z ₁ , z ₂ ,..., z _N }; the centroid of the template image obtained after normalization and smoothing The height increment feature matrix is denoted as F(Z): Among them, 1<r≤N; is the rth feature vector of the template image; the centroid height incremental feature matrix of the target image obtained after normalization and smoothing is denoted as F(M): /> Among them, 1<e<N;/> is the e-th feature vector of the target image; the first feature distance is defined according to the DTW algorithm as:

in, Use the following recursive formula to convert to the solution of F(M) and F(Z) subsequence problems:

The first characteristic distance between the target image contour sampling point set M and the template image contour sampling point set Z is: Dis(M, Z)=d _DTW (F(M), F(Z)).

Further, by using the DTW algorithm to compare the centroid height incremental feature matrix of the template image and the flipped image, obtaining the second feature distance specifically includes: flipping the target image to obtain a flipped image; setting the flipped image contour sampling point set as M _x ; The centroid height incremental feature matrix of the target image is denoted as F(M _x ); the second feature distance between the flipping image contour sampling point set M _x and the template image contour sampling point set Z is: Dis(M _x , Z)= d _DTW (F(M _x ), F(Z)); Calculating the minimum characteristic distance between the first characteristic distance and the second characteristic distance through the minimum value function specifically includes: using the following formula to calculate the minimum characteristic distance: D(M, Z)=min(Dis(M,Z), Dis( _Mx ,Z)).

Further, introducing the shape complexity into the minimum feature distance to obtain the similarity distance specifically includes: defining the shape complexity as: Among them, std represents the standard deviation; after introducing the shape complexity, the following formula is used to obtain the similarity distance:

Further, the contour target recognition method based on the centroid height increment combined with DTW also includes the following steps: establishing a contour template library, the contour template library includes a plurality of template images; Matching, output the template image that matches the target image as the recognition result.

Further, the contour template library includes the original image of the template image and the geometrically transformed image of the template image; the geometrically transformed image is obtained by translating, rotating, and scaling the original image.

Compared with the prior art, the present invention uses a contour target recognition method combining centroid height increments and DTW for target recognition. When the DTW algorithm compares the distance values of the centroid height increment matrix of the target image and the template image, only The cumulative distance of the previous column does not need to keep all the data, thereby reducing the time complexity of the algorithm, which is more conducive to improving the recognition processing speed, and is suitable for real-time target recognition; the present invention also flips the target image, and utilizes the DTW algorithm to The flipped image and the template image are re-identified to improve the accuracy of target recognition and avoid recognition errors caused by image flipping, and because the recognition processing speed of the DTW algorithm is faster, although the second recognition is performed, the overall operation speed remains the same. Soon, the accuracy and calculation speed are well balanced; the present invention also introduces shape complexity to further improve the contour matching effect. The higher the shape complexity, the lower the sensitivity to local deformation of the contour, and the better the recognition result It is credible, which is conducive to improving the noise immunity of target recognition.

Description of drawings

The above and other objects, features and advantages of exemplary embodiments of the present invention will become readily understood by reading the following detailed description with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary rather than restrictive manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

Fig. 1 schematically shows a schematic flowchart of a contour target recognition method based on a combination of centroid height increment and DTW provided by an optional embodiment of the present invention;

Fig. 2 schematically shows a flow chart of comparing a target image with a template image in the present invention;

Fig. 3 schematically shows a schematic flow diagram of a part of the process after introducing shape complexity in the present invention;

Fig. 4 schematically shows a schematic diagram of extracting the peripheral contour of an image and extracting a plurality of sampling points on the peripheral contour to generate a contour sampling point set in the present invention;

Fig. 5 schematically shows feature maps of different sampling points of different shapes, the feature maps are formed according to the smoothed feature sequence C _i of the corresponding sampling points;

Fig. 6 schematically shows a schematic diagram of target recognition results.

Detailed ways

Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be embodied in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided for more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

It should be noted that, unless otherwise specified, the technical terms or scientific terms used in the present invention shall have the usual meanings understood by those skilled in the art to which the present invention belongs. In this document, relational terms such as "first" and "second", etc., are only used to distinguish one entity or operation from another, and do not necessarily require or imply that there is a relationship between these entities or operations. There is no such actual relationship or sequence. Terms such as "connection" and "connection" should be understood in a broad sense, for example, it can be fixed connection, detachable connection, or integrated; it can be mechanical connection or electrical connection; it can be direct connection or It can be connected indirectly through an intermediary. The term "comprises", "comprises" or any other variation thereof is intended to cover a non-exclusive inclusion such that a process, method, article or apparatus comprising a set of elements includes not only those elements but also other elements not expressly listed elements, or also elements inherent in such a process, method, article, or apparatus. Without further limitations, an element defined by the statement "comprising..." does not exclude the presence of additional same elements in the process, method, article or device comprising said element.

The invention provides a contour target recognition method based on the combination of the centroid height increment and DTW, comprising the following steps: obtaining the centroid height increment feature matrix of the template image and the target image; Compare the centroid height incremental feature matrix to obtain the first feature distance; after flipping the target image, obtain the centroid height incremental feature matrix of the flipped image; use the DTW algorithm to compare the centroid height incremental feature matrix of the template image and the flipped image Perform comparison to obtain the second feature distance; use the minimum value function to find the minimum feature distance between the first feature distance and the second feature distance; introduce shape complexity into the minimum feature distance to obtain the similarity distance, according to the similarity distance Output the recognition result.

The present invention adopts DTW algorithm to carry out target recognition, and DTW algorithm only retains the cumulative distance of the previous column when budgeting in the centroid height incremental matrix, and does not need to retain all data, thereby reducing the time complexity of the algorithm, thereby more conducive to improving Recognition processing speed, suitable for real-time target recognition; the invention also flips the target image, and uses the DTW algorithm to perform secondary recognition on the flipped image and the template image, thereby improving the accuracy of target recognition and avoiding recognition caused by image flipping error, and because the recognition processing speed of the DTW algorithm is fast, although the secondary recognition is performed, the overall calculation speed is still very fast, which balances the accuracy and calculation speed well; the present invention also introduces shape complexity to further improve the outline Matching effect, the higher the complexity of the shape, the lower the sensitivity of the local deformation of the contour, so the recognition results are more reliable, which is conducive to improving the noise resistance of target recognition.

Optionally, obtaining the centroid height incremental feature matrix includes the following steps: preprocessing the image; extracting the peripheral contour of the image; extracting a plurality of sampling points on the peripheral contour to generate a contour sampling point set; calculating the contour sampling point set Contour centroid; calculate centroid height; calculate centroid height increment; find centroid height increment sequence; find centroid height increment matrix.

Optionally, preprocessing the image specifically includes the following steps: converting the image from a three-channel image to a single-channel image; performing thresholding noise reduction processing on the single-channel image; extracting the peripheral contour of the image includes the following steps: using Canny differential algorithm Edge extraction is performed on the image; image morphology operations are used to expand the thin edges and fill holes to form a complete peripheral contour; again, the Canny differential operator is used to accurately extract the peripheral contour.

Optionally, extracting a plurality of sampling points on the peripheral contour, and generating a contour sampling point set specifically includes: extracting N sampling points at equal intervals on the peripheral contour, and generating a contour sampling point set M; wherein, M={m _j }= {m _i , m ₂ , . . . , m _N }.

Optionally, calculating the contour centroid of the contour sampling point set specifically includes: calculating the contour centroid Q(x ₀ , y ₀ ) of the contour sampling point set using the following formula:

Optionally, calculating the centroid height specifically includes: defining the Euclidean distance between any sampling point m _j (x _j , y _j ) and the contour centroid Q(x ₀ , y ₀ ) as the centroid height of any sampling point m _j , j=(1, 2,...N); use the following formula to calculate the centroid height g _j between any sampling point m _j (x _j , y _j ) and the contour centroid Q(x ₀ , y ₀ ):

Optionally, the calculation of the centroid height increment specifically includes: defining the difference between the centroid height g _j of any sampling point m _j and the centroid height g _i of a certain sampling point m _i , which is relative to any sampling point m _j The centroid height increment h _{i, j} at a certain sampling point m _i , i=(1, 2, ... N); obtaining the centroid height increment sequence specifically includes: calculating all sampling points relative to a certain sampling point The centroid height increment of m _i , and the centroid height increments of all sampling points relative to a certain sampling point m _i are arranged in the _order of sampling points, and the centroid height sequence H _i of sampling point mi can be obtained: Hi=(h _{i , i} , h _{i, i+1} , h _{i, N} , h _{i, 1} ,...h _{i, i+1} ) ^T ; the sequence includes N elements, corresponding to N sampling points relative to sampling point m The centroid height increment for _i .

Optionally, obtaining the centroid height increment matrix specifically includes: arranging the centroid height increment sequence H _i corresponding to all sampling points according to the order of the sampling points to obtain the centroid height increment matrix L(M): L(M)=( H ₁ , H ₂ , . . . , H _N-1 , H _N ).

Optionally, obtaining the feature matrix of the centroid height increment further includes the following steps: performing normalization processing on the centroid height increment h _{i, j} to obtain the normalized centroid height increment ||h _{i, j} || is the modulus of the centroid height increment; the normalized centroid height increment sequence In this way, the centroid height increment after normalization /> Has scaling invariance.

Add a positive integer coefficient d to the centroid height increment sequence _Hi after normalization processing, d (1<d<N); divide the centroid height increment sequence Hi after normalization processing into S disjoint sub- Sequence [1,d], [1+d,2d]..., wherein, S=[N/d]; use the following formula to calculate the average value c _i,t of the centroid height increment of each subsequence:

Among them, t=1, 2,...S; f=(t-1)×d, f is the centroid height increment of d sequences, and the centroid height increment after smoothing is the centroid height increment of d sequences Calculate the average value, and after adding d, the feature matrix of the image is smoothed and reduced in dimension.

Arrange the S mean value data in an orderly manner to obtain the feature sequence C _i of the sampling point m _i after smoothing processing: C _i =(ci _,1 , _ci,2 ,...,ci _,S-1 , ci _{, S} ) ^T ; the centroid height incremental feature sequence C _i after smoothing of all sampling points is arranged in the order of sampling points, and the smoothed centroid height incremental feature matrix F(M) is obtained: F(M) =(C ₁ , C ₂ , . . . , C _N-1 , C _N ). After smoothing, it will not be too sensitive to the local deformation of the contour of the noise, the feature dimension is reduced, the calculation is simpler, and a good compromise has been achieved between accuracy, noise resistance, and simplicity. After smoothing, not only the robustness of the descriptor to contour deformation and noise interference is improved, but also the dimension of the feature vector is reduced to facilitate subsequent matching.

Optionally, using the DTW algorithm to compare the centroid height incremental feature matrix of the template image and the target image, and obtaining the first feature distance specifically includes: setting the sample point set of the target image contour as M, M={m ₁ , m ₂ ,...,m _N }; set the template image contour sampling point set as Z, Z={z ₁ ,z ₂ ,...,z _N }; the template image obtained after normalization and smoothing The centroid height incremental feature matrix is denoted as F(Z): Among them, 1<r≤N;/> is the rth feature vector of the template image; the centroid height incremental feature matrix of the target image obtained after normalization and smoothing is denoted as F(M): /> Among them, 1<e<N;/> is the e-th feature vector of the target image; the first feature distance is defined according to the DTW algorithm as:

in, Use the following recursive formula to convert to the solution of F(M) and F(Z) subsequence problems:

The first characteristic distance between the target image contour sampling point set M and the template image contour sampling point set Z is: Dis(M, Z)=d _DTW (F(M), F(Z)). In this way, the DTW algorithm is used to compare the target image and the template image, and judge whether the target image and the template image match according to the size of the first feature distance. The smaller the first feature distance, the more the target image matches the template image.

Among them, C ₁ to C _N in F(M) are The two-dimensional matrix of , when using the DTW algorithm for comparison, it is necessary to compare the point features in the matrix. Therefore, F(M)=(C ₁ , C ₂ ,..., C _N - ₁ , C _N ) expressed as

Optionally, by using the DTW algorithm to compare the centroid height incremental feature matrix of the template image and the flipped image, obtaining the second feature distance specifically includes: flipping the target image to obtain a flipped image; setting the flipped image contour sampling point set as M _x ; The centroid height increment feature matrix of the target image is denoted as F(M _x ); The second feature distance between the flipping image contour sampling point set M _x and the template image contour sampling point set Z is: Dis(M _x , Z) =d _DTW (F(M _x ), F(Z)). In this way, use the DTW algorithm to compare the flipped image of the target image with the template image, and judge whether the flipped image of the target image matches the template image according to the size of the second feature distance. The smaller the second feature distance, the smaller the flipped image of the target image The closer it matches the template image.

Optionally, calculating the minimum characteristic distance between the first characteristic distance and the second characteristic distance through the minimum value function specifically includes: calculating the minimum characteristic distance by using the following formula:

D(M,Z)=min(Dis(M,Z), Dis(M _x ,Z)).

In this way, it is judged whether the target to be tested matches the template image according to the minimum feature distance, and the smaller the minimum feature distance is, the more the target to be tested matches the template image. Matching the flip image of the target to be tested with the template image twice can improve the matching accuracy.

Optionally, the flipped image of the target to be measured can be obtained by flipping the target image.

Optionally, introducing the shape complexity into the minimum feature distance to obtain the similarity distance specifically includes: defining the shape complexity as: Among them, std represents the standard deviation; after introducing the shape complexity, the following formula is used to obtain the similarity distance:

In this way, the introduction of shape complexity can improve the noise resistance of the contour target recognition method and further improve the matching effect. After the shape complexity is introduced, the higher the shape complexity, the lower the sensitivity to local deformation of the contour, and the more reliable the recognition result is.

Optionally, the contour target recognition method based on the centroid height increment combined with DTW also includes the following steps: setting up a contour template library, which includes a plurality of template images; combining the target image with a plurality of template images in the contour template library Matching is performed, and the template image matched with the target image is output as the recognition result. Wherein, the template image is an image of a certain object, and the target image is an image of an uncertain object, and the object type of the target image can be identified through the outline object recognition method of the present application.

Optionally, the contour template library includes the original image of the template image and the geometrically transformed image of the template image; the geometrically transformed image is obtained by translating, rotating, and scaling the original image. In this way, the technical solution provided by the present application has translation, rotation, and scaling invariance.

An optional specific embodiment of the present application will be described in detail below with reference to FIG. 1 to FIG. 6 .

As shown in Figure 1, the contour target recognition method based on the combination of centroid height increment and DTW in the optional specific embodiment provided by the present application includes the following steps:

(1) Image preprocessing and contour extraction: Preprocess the target image and several template images and extract their respective peripheral contours with contour extraction algorithms. Wherein, the target to be tested is an image of an uncertain object, and the template image is an image of a certain object.

(2) Perform uniform sampling and centroid extraction on the contour: extract the outer contour of the target to be measured, select N sampling points on the outer contour, and extract the contour centroid of the N sampling points. With the contour centroid as a reference point, the centroid height increment descriptor is established according to the height relationship of other points compared with this point: for the selected sampling points, their respective distances from the centroid and the centroid height increment are calculated.

(3) Normalize and smooth the centroid height incremental feature: reduce the dimensionality of the centroid height incremental feature obtained in step (2) and make it invariant to translation and rotation.

(4) Use the DTW algorithm to calculate the feature distance of two contours: use the processed centroid height incremental feature obtained in step (3) to calculate the matching cost and then calculate the feature distance between the contours.

(5) Flip the target image and compare it with the template library image twice, and apply the minimum value function to find the minimum similarity distance to obtain the initial recognition result, and then combine the shape complexity analysis to recognize the target image again, and get The final recognition result.

Wherein step (1) includes the following steps (1.1) to (1.5):

(1.1) The image is converted from a three-channel image to a single-channel image;

(1.2) Thresholding and denoising the image;

(1.3) using Canny differential operator to extract the edge of the image;

(1.4) Use image morphological operations to expand the thin edges and fill the holes to form a complete peripheral outline;

(1.5) The Canny differential operator is used again to accurately extract the peripheral contour.

The step (2) includes the following steps (2.1) to (2.2):

(2.1) Extract N valid feature points on the image outer contour;

(2.2) Calculate the contour centroid according to the algorithm;

(2.3) Find the distance between the sampling point and the center of mass, i.e. the height of the center of mass;

(2.4) Obtaining the centroid height increment sequence;

(2.5) Finding of centroid height increment matrix;

As shown in Figure 4(a)(b), the extraction of contour feature points in step (2.1) is sampling at equal intervals, and a certain number of contour points are selected. Including the following steps: extracting the complete contour point set of the image contour, selecting the number of contour points as N, dividing the total contour point m by the required number of points N, the sampling distance m/N of the contour can be obtained, and selecting the contour point The starting point can be random. From this, it can be concluded that the larger the number of sampling points N is, the more accurate the described shape will be.

As shown in Figure 4(c), step (2.2) utilizes step (2.1) to obtain the point set of the contour part as M={m _i }={m ₁ , m ₂ ,..., m _N }, where N is the sampling point The number of , m _i is the i-th sampling point of the contour. Assuming the sampling point m _i ( _xi , y _i ), the calculation formula of the contour centroid Q(x ₀ , y ₀ ) is as follows:

Step (2.3) is to calculate the centroid height increment of the contour. The most important thing is to find the distance between the sampling point and the centroid, that is, the centroid height value, which represents the position of the sampling point on the contour. For the sampling point m _i (x _i , y _i ), define the centroid height g _i as the Euclidean distance between the point and the centroid Q(x ₀ , y ₀ ), that is:

As shown in Figure 5, the marked contour and the feature map of different sampling points A, B, and C of the contour are given. The feature map is formed according to the smoothed feature sequence C _i of the corresponding sampling points, from which it can be verified that the same The difference of different sampling points of the contour.

In step (2.4) to obtain the centroid height increment sequence, the centroid height increments of all sampling points relative to point _mi are arranged in the order of contour points, and the centroid height increment sequence H _i of sampling point _mi is obtained, namely:

H _i = (h _{i, i} , h _{i, i+1} , .. h _{i, N} , h _{i, 1} , .. h _{i, i-1} ) ^T ;

As shown in Figure 4, step (2.5) calculates the centroid height increment matrix, arranges the centroid height increment sequence H _i corresponding to each point on the shape contour M according to the order of the contour points, and obtains a matrix with a size of N′N :

L(M)=(H ₁ , H ₂ , . . . , H _N-1 , H _N );

where L(M) represents the centroid height increment matrix of the contour, and the i-th column of the matrix represents the centroid height increment descriptor of the sampling point m _i on the contour M. This descriptor describes the relative height relationship between contour points and points, which does not change with the rotation and translation of the contour.

In step (3), normalizing and smoothing the centroid height incremental feature includes the following steps:

To make this descriptor scale-invariant, we normalize each row of the matrix:

where ||h _{t, j} || is the modulus of the centroid height incremental data. This formula defines the centroid height increment of the sampling point _mi relative to all sampling points of the contour, which effectively describes the contour information, but is too sensitive to the local deformation of the contour caused by noise, and the feature dimension is too high and the calculation is complicated. Here, a smooth strategy is adopted to achieve a good compromise between the accuracy, noise resistance, and simplicity of the descriptor. The specific process is as follows:

The centroid height increment of point _mi can be expressed as:

The sequence contains N elements, corresponding to the centroid height increment of N sampling points relative to the point, adding a positive integer coefficient d (1<d<N), dividing the sequence into S disjoint subsequences [1 , d], [1+d, 2d]..., where S=[N/d], calculate the average of the centroid height increments for each sequence:

In the formula, t=1, 2, ... S. Arrange the S mean value data in an orderly manner to obtain the feature sequence C _i of point m _i after smoothing processing, namely: C _i =(ci _,1 , _ci,2 ,...,ci _{,S- 1} , c _{i, s} ) ^T ;

After smoothing, not only the robustness of the descriptor to contour deformation and noise interference is improved, but also the dimension of the feature vector is reduced to facilitate subsequent matching. Arrange the smoothed descriptors of all sampling points in order to obtain the centroid height incremental feature matrix F(M) of the contour M:

F(M)=(C ₁ , C ₂ , . . . , C _N-1 , C _N );

Step (4) adopting the DTW algorithm to calculate the feature distance of the two contours includes the following steps:

As shown in Figure 2 and Figure 3, after the feature descriptor of the shape is obtained in step (2), the similarity between the two shapes is calculated. Since the centroid height increment descriptor contains the global feature of the contour point set order, this paper selects The DTW algorithm matches the obtained shape features.

The DTW algorithm adopts the idea of dynamic programming, and transforms the problem of finding the similarity between the target contour and the template contour into the shortest path planning problem between the two contour points. The more similar the features between different contour sampling points, the smaller the corresponding path distance. Set the target contour point set M={m ₁ ,m ₂ ,...,m _N } and the template image contour point set Z={z ₁ ,z ₂ ,...,z _N }, extract the centroid height Incremental feature, get centroid height incremental matrix F(M) and F(Z). Eigenmatrix of hypothesized object contours

Among them, 1<e<N;, the feature matrix of the template image where, 1<r≤N; where /> and /> represent the e-th feature vector of the target image and the r-th feature vector of the template image, respectively. Then the similarity distance difference between two shapes is defined according to the DTW algorithm as:

In the formula

The DTW algorithm adopts the idea of dynamic programming, and uses the following recursive formula to transform the above problem into the solution of the F(M) and F(Z) subsequence problems:

That is, the similarity distance between contours M and Z is:

Dis(M,Z)=d _DTW (F(M),F(Z));

In step (5), specifically comprise the following steps:

(5.1) Flip the target image and the template image for secondary matching, and apply the minimum value function to obtain the minimum distance between contours in the two comparisons.

As shown in Fig. 2, in the step (5.1), the target to be recognized often appears to be flipped during the image recognition process, which increases the probability of false matching. Therefore, the contour M of the image to be recognized is flipped to obtain M _x , and the shape M _x and shape Z are matched to obtain the distance between them:

Dis(M _x , Z)=d _DTW (F(M _x ), F(Z))

(5.2) Combined shape complexity analysis to get the final recognition result.

As shown in Figure 6, in step (5.2), the higher the complexity of the shape, the lower the sensitivity to the local deformation of the contour, and the more reliable the recognition result is, so the introduction of shape complexity further improves the matching effect of the contour , the complexity of defining the shape profile is:

where std stands for standard deviation. By introducing shape complexity, the similarity distance Q(M, Z) between two shapes is finally obtained:

where Q(M) and Q(Z) are the complexity of shapes M and Z, respectively.

Aiming at the target recognition problem based on contour features, this application proposes a contour target recognition method based on the combination of centroid height increment and DTW, which combines the centroid height increment feature descriptor with the DTW similarity measurement algorithm. Extract the sampling points, and extract the centroid height incremental features of the target image and the contour points of the template image, and then use the DTW algorithm to find a regular path to measure the similarity of the feature matrix of the target image and the template image, and finally combine the shape of the contour Complexity analysis obtains the final recognition result. This method can ensure that the recognition rate is better than most common traditional target recognition algorithms, and at the same time improve the real-time performance of target recognition.

The contour target recognition method based on the combination of centroid height increment and DTW provided by the preferred embodiment of the present application includes the following steps: establishing a contour template library of multiple types of target images and template images; performing image preprocessing and The peripheral contour is extracted, and a contour point set is generated; the centroid height increment feature of the contour is extracted from the template image and the target image respectively, and normalized and smoothed; the centroid height increment of the template image and the target image is calculated by the DTW algorithm. Measure the feature distance; flip the target image and compare it with the template library image twice, and use the minimum value function to find the minimum similarity distance to obtain the initial recognition result, and then combine the shape complexity analysis to recognize the target image again, and get the final recognition result.

Preferably, the contour template library specifically includes geometric transformation features and complete image features of the image.

Preferably, image preprocessing and peripheral contour extraction specifically include the following steps: the image is converted from a three-channel image to a single-channel image; the image is thresholded for noise reduction; the Canny differential operator is used to extract the edge of the image; The operation expands the thin edge and fills the hole to form a complete peripheral contour; again, the Canny differential operator is used to accurately extract the peripheral contour.

Preferably, extracting the centroid height incremental feature of the contour, and performing normalization processing and smoothing processing specifically includes the following steps: obtaining the contour centroid; extracting the centroid height incremental feature of the target image through a feature extraction algorithm: finding N The centroid height of each point is obtained, and then the centroid height increment of N points is calculated, and the centroid height increment of any sampling point is arranged in contour sequence to obtain the centroid height increment sequence, and then the centroid height increment sequence of N points is obtained. The centroid height increment matrix; normalize the centroid height increment feature of the contour sampling points; perform smoothing processing on the centroid height increment features of the contour sampling points.

Preferably, calculating the centroid height incremental feature distance between the template image and the target image through the DTW algorithm specifically includes the following steps: using the DTW algorithm to quantify the similarity of the centroid height incremental feature between the sampling points of the target contour and the template contour. The DTW algorithm transforms the general problem of obtaining the similarity between two contours into a sub-problem of obtaining the similarity between contour points.

The contour target recognition method based on centroid height increment and DTW provided by the preferred embodiment of the present invention belongs to the technical field of machine vision and target detection, and solves the problem of how to balance the speed and accuracy of recognition in the prior art due to the low speed of recognition processing . Perform preprocessing and contour extraction on the template image and the image to be recognized, uniformly sample the contour and calculate the centroid, use the contour centroid as a reference point, and construct a centroid height incremental descriptor according to the height relationship of other sampling points relative to this point, Normalize and smooth the centroid height incremental features; use the DTW algorithm to obtain the feature distance of the two contours, then flip the target image for secondary recognition, define the minimum function to obtain the shortest similarity distance in the two recognitions , and finally combined with the complexity analysis of contour features to screen out the final recognition results. The present invention combines the contour description method of centroid height increments with the DTW algorithm, which can ensure a good recognition rate effect and improve the real-time performance of target recognition.

The beneficial technical effects of the present invention at least include: in the case of low recognition accuracy and high complexity of the traditional contour target algorithm, the present invention considers combining the centroid height incremental feature descriptor with spatial relationship with the DTW algorithm, during the recognition process The complexity of the algorithm is reduced by smoothly reducing the feature dimension. In the case of low target recognition accuracy, consider flipping the target image for secondary recognition and combine shape complexity to improve target shape recognition accuracy and anti-noise.

It is verified by experiments that the retrieval rate of the present application can reach more than 90%, which is higher than the existing contour recognition algorithm. In addition, the total time complexity of the present application is O(N ² )+O(N ² ), which is significantly lower than the time complexity O(N ² )+O(N ³ ) of the existing contour recognition algorithm. The technical solution provided by this application can effectively improve the real-time performance of target recognition while ensuring a good retrieval rate, balance efficiency and accuracy well, and have better overall performance.

The above is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Anyone skilled in the art can easily think of changes or substitutions within the technical scope disclosed in the present invention. Should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the protection scope of the claims.

Claims (10)

1. The contour target identification method based on the combination of centroid height increment and DTW is characterized by comprising the following steps:

obtaining a centroid height increment feature matrix of the template image and the target image;

comparing the centroid height increment feature matrix of the template image with that of the target image through a DTW algorithm to obtain a first feature distance;

after the target image is overturned, a centroid height increment feature matrix of the overturned image is obtained;

comparing the centroid height increment feature matrix of the template image with that of the overturning image through a DTW algorithm to obtain a second feature distance;

obtaining the minimum feature distance from the first feature distance and the second feature distance through a minimum function;

and introducing shape complexity into the minimum feature distance to obtain a similarity distance, and outputting a recognition result according to the similarity distance.

2. The method for identifying a contour target based on combination of a centroid height increment and DTW according to claim 1, wherein the step of obtaining a centroid height increment feature matrix comprises the steps of:

preprocessing an image;

extracting the peripheral outline of the image;

extracting a plurality of sampling points on the peripheral contour to generate a contour sampling point set;

calculating the contour centroid of the contour sampling point set;

calculating the height of the mass center;

calculating the height increment of the mass center;

solving a centroid height increment sequence;

and (5) solving a centroid height increment matrix.

3. The method for identifying a contour object based on a combination of a centroid height increment and DTW as recited in claim 2, wherein,

the preprocessing of the image specifically comprises the following steps: converting the image from a three-channel image to a single-channel image; thresholding and noise reduction treatment is carried out on the single-channel image;

the extracting the peripheral outline of the image comprises the following steps: carrying out edge extraction on the image by adopting a Canny differential operator; adopting image morphology to operate and expand the tiny edges and filling the holes to form a complete peripheral outline; and accurately extracting the peripheral outline by adopting a Canny differential operator again.

4. The method for identifying a contour object based on a combination of a centroid height increment and DTW as recited in claim 2, wherein,

extracting a plurality of sampling points on the peripheral contour, and generating a contour sampling point set specifically comprises: extracting N sampling points on the peripheral contour at equal intervals to generate a contour sampling point set M; wherein M= { M _j }＝{m ₁ ，m ₂ ，...，m _N }；

The calculating the contour centroid of the contour sampling point set specifically comprises the following steps: the centroid Q (x) of the profile sample point set is calculated using the following formula ₀ ，y ₀ )：

Calculating the centroid height specifically includes: defining any sampling point m _j (x _j ，y _j ) And the centroid Q (x) ₀ ，y ₀ ) The Euclidean distance between the two is any sampling point m _j Centroid height of j= (1, 2,..n); any sampling point m is calculated by adopting the following formula _j (x _j ，y _j ) And the centroid Q (x) ₀ ，y ₀ ) Centroid height g between _j ：

Calculating the centroid height increment specifically includes: defining any sampling point m _j Centroid height g _j And a certain sampling point m _i Centroid height g _i The difference between the two is any sampling point m _j With respect to a certain sampling point m _i Centroid height increment h _i，j ，i＝(1，2，...N)；

The sequence for solving the centroid height increment specifically comprises the following steps: calculating all sampling points relative to a certain pointSampling point m _i And increasing the height of the centroid of the image data set (C) by comparing all the sampling points with a certain sampling point m _i The centroid height increment of (2) is arranged according to the sequence of sampling points to obtain a sampling point m _i Centroid height sequence H _i ：

H _i ＝(h _i，i ，h _i，i+1 ，h _i，N ，h _i，1 ，...h _i，i+1 ) ^T The method comprises the steps of carrying out a first treatment on the surface of the The sequence includes N elements corresponding to N sampling points relative to sampling point m _i Is a centroid height increment of (2);

the calculating of the centroid height increment matrix specifically comprises: centroid height increment sequence H corresponding to all sampling points _i And (3) arranging according to the sequence of sampling points to obtain a centroid height increment matrix L (M):

L(M)＝(H ₁ ，H ₂ ，...，H _N-1 ，H _N )。

5. the method for identifying a contour target based on a combination of a centroid height increment and DTW as recited in claim 4, wherein the step of obtaining a centroid height increment feature matrix further comprises the steps of:

for centroid height increment h _i,j Normalization processing is carried out to obtain centroid height increment after normalization processing||h _i，j The I is a module of the centroid height increment;

normalized centroid height increment sequence

For the centroid height increment sequence H after normalization processing _i Adding a positive integer coefficient d, d (d is more than 1 and less than N);

the centroid height increment sequence H after normalization processing _i Divided into S disjoint sub-sequences [1, d]、[1+d，2d].., wherein s= [ N/d ]]；

Calculating centroid height increment of each sub-sequence by using the following formulaAverage value c of (2) _i,t ：

Wherein t=1, 2,..s;

s average value data are orderly arranged to obtain a sampling point m _i Feature sequence C after smoothing _i ：C _i ＝(c _i，1 ，c _i,2 ，...，c _i,S-1 ，c _i,S ) ^T ；

Centroid height increment feature sequence C after smoothing all sampling points _i And (3) arranging according to the sequence of sampling points to obtain a smoothed mass center height increment feature matrix F (M):

F(M)＝(C ₁ ，C ₂ ，...，C _N-1 ，C _N )。

6. the method for contour object recognition based on a combination of centroid height increment and DTW as recited in claim 5, wherein,

comparing the centroid height increment feature matrix of the template image and the target image through a DTW algorithm, wherein the obtaining of the first feature distance specifically comprises the following steps:

let the target image contour sampling point set be M, M= { M ₁ ，m ₂ ，...，m _N }；

Let the template image contour sampling point set be Z, Z= { Z ₁ ，z ₂ ，...，z _N }；

The centroid height increment feature matrix of the template image obtained after normalization processing and smoothing processing is marked as F (Z):

wherein r is more than 1 and less than or equal to N;the r characteristic vector of the template image;

the centroid height increment feature matrix of the target image obtained after normalization processing and smoothing processing is marked as F (M):

wherein e is more than 1 and less than N;an e-th feature vector of the target image;

defining the first feature distance as follows according to the DTW algorithm:

wherein ,

the following recursive formula is used to translate into a solution to the F (M) and F (Z) subsequence problem:

the first feature distance between the target image contour sampling point set M and the template image contour sampling point set Z is as follows:

Dis(M，Z)＝d _DTW (F(M)，F(Z))。

7. the method for contour object recognition as defined in claim 6, wherein,

comparing the centroid height increment feature matrix of the template image and the overturning image through a DTW algorithm, wherein the obtaining of the second feature distance specifically comprises the following steps:

turning over the target image to obtain a turned-over image;

let the sampling point set of the image contour of the turnover be M _x ；

The centroid height increment feature matrix of the target image is denoted as F (M _x )；

Turnover image contour sampling point set M _x And the second characteristic distance between the template image contour sampling point set Z is as follows:

Dis(M _x ，Z)＝d _DTW (F(M _x )，F(Z))；

the obtaining the minimum feature distance between the first feature distance and the second feature distance through the minimum function specifically comprises the following steps: the minimum feature distance is calculated using the following formula:

D(M，Z)＝min(Dis(M，Z)，Dis(M _x ，Z))。

8. the method for contour object recognition based on a combination of a centroid height increment and DTW as recited in claim 7, wherein,

the step of introducing shape complexity into the minimum feature distance to obtain the similarity distance specifically comprises the following steps:

the shape complexity is defined as:

wherein std represents standard deviation;

after the shape complexity is introduced, the similarity distance is calculated by adopting the following formula:

9. the profile-target recognition method based on the combination of the centroid height increment and the DTW according to any one of claims 1 to 8, further comprising the steps of:

establishing a profile template library, wherein the profile template library comprises a plurality of template images;

and matching the target image with a plurality of template images in the outline template library, and outputting the template image matched with the target image as a recognition result.

10. The method for identifying a contour target based on the combination of centroid height increment and DTW according to claim 9, wherein the contour template library comprises an original image of a template image and a geometrically transformed image of the template image; and obtaining the geometric transformation image by translating, rotating and scaling the original image.