CN103093461A - Shape matching method based on measurement information - Google Patents

Abstract

The invention belongs to the field of computer vision and relates to a shape matching method based on measurement information. Firstly, the measurement information of shape outline points are used as a feature descriptor and a restricted relationship between the outline points and the whole shape is established. Secondly, in order to increase adaptability of shape local features of the shape matching method based on the measurement information to micro-deformation, the measurement information is divided into subsections and is smoothened. Finally, the shape matching method based on the measurement information is popularized and a feature description sign is generated by means of the measurement information such as Euclidean distance and the triangular radius. The shape matching method based on the measurement information is characterized in that the shape feature description sign is relatively simple. The shape matching method based on the measurement information can conduct description and matching operation to plane target characteristics only simple section modeling and data calculation are needed, and therefore calculation time is greatly shortened and generality of the shape matching method is improved. The shape matching method based on the measurement information has the advantages of being free from changing characteristics in the process of translation and rotation, simple in calculation, low in dimensionality and stable in local deformation.

Description

A kind of shape matching method based on metric

Technical field

The invention belongs to computer vision field, relate to shape description and coupling, specially refer to a kind of shape matching method based on metric.

Background technology

Form fit is an important research direction in computer vision, contacts closely with directions such as object identification, character recognition, image retrieval, medical image analysis.Shape just receives publicity as the important evidence of mankind's observation and recognition object all the time.And the purpose of form fit is sought a kind of simple, effective shape facility describing mode exactly, can distinguish the shape difference of variety classes object, can tolerate again the shape distortion of Different Individual in similar object.So key is exactly to find a suitable equilibrium point, obtain optimum efficiency between the property distinguished and deformability.

At present more common shape facility describing method mainly is divided into based on the method for global information with based on the method for local message.Shape description method based on global information is mainly paid close attention to the shape global feature, and is insensitive for shape local deformation, but also therefore can't describe the minutia of shape, and common method comprises Fourier transform, the methods such as wavelet transformation.On the contrary, shape description method based on local message is mainly paid close attention to the local feature of shape, usually each point of shape there is description accurately, but noise and deformation comparison for the part are responsive, common method comprises curvature scale space, Shape context (Shape Context, SC) etc.For the complementarity of global approach and partial approach, method is arranged with both effective combinations, Shape tree has been proposed, the describing methods such as Contour flexibility, and obtained good effect.Yet this compound method characteristic descriptor generates complicated, and computation complexity is high, and form fit efficient is low, is unfavorable for using in practice.

The Chinese invention patent of Harbin Engineering University, application number 201210150026, " based on the shape description method of fractional Fourier transform ", this patent is based on fractional Fourier transform, the method has guaranteed to make described feature both relevant with the boundary profile of object, relevant with the interior zone of object again, yet to the description a bit deficient in of the minutia of shape.

The Chinese invention patent of Tongji University, application number 200910046267, " based on the image one-way point-to-point matching method of minimum deformation energy ", this patent is based on the Patch-based match method, and calculated amount is smaller, yet the method needs to specify in advance initial match point, a bit deficient in aspect the self-adaptation of method, and the method is also inadequate to the discrimination of local microdeformation.

Summary of the invention

The present invention proposes a kind of new shape description and matching process based on metric, solved the deficiencies in the prior art, the shape facility descriptor that the method generates is relatively simple, only needs metric to be described.

Technical scheme of the present invention comprises the steps:

Step 1. pair original image utilizes the method for rim detection to extract profile.

Step 2. is carried out uniform sampling on profile, sampling number is N.N configuration sampling point is expressed as X={x ₁, x ₂..., x _N.

Step 3. is calculated arbitrary configuration sampling point x _i(i=1,2 ..., metric N).

Step 3-1 is with x _iBe starting point, scan counterclockwise other configuration sampling points, calculate x _iEuclidean distance (Euclidean Distance, ED) to other all configuration sampling points.

Step 3-2 is from x _iPoint is done the tangent line of profile, and in the tangent line left side, distance is being for just when other configuration sampling point, and distance is to bear on the tangent line right side.

The x that step 3-3 will calculate above _iPoint forms following proper vector D to the distance of other configuration sampling points _i

D _i=(d _i，i,d _i，i+1,...,d _i,N,d _i，1,...,d _i，i-1) （1）

Wherein, d _i,jExpression x _iTo x _jDistance, d _{I, i}Value be 0.

Can be with leg-of-mutton area as metric to the calculating of metric in step 3, namely can be in order to lower step replacement step 3.Concrete operation method is as follows:

Step 3-1. is for the sampled point x on profile _i, some x _i-1And x _i+1Respectively its left and right consecutive point, other arbitrary point x on profile _jCan both and x _i-1And x _i+1Consist of a triangle.If d ₁Expression x _i-1And x _jBetween Euclidean distance, d ₂Expression x _i+1And x _jBetween distance, d ₃Expression x _i-1And x _i+1Between distance.

Step 3-2. calculates leg-of-mutton semi-perimeter L with following formula _{I, j}, area H _i,jAnd radius R _{I, j}:

$L_{i,j}=\frac{1}{2}({\mathrm{<figure-callout\; id="1"\; label="d"\; filenames="HDA00002731082100032.png"\; state="\{\{state\}\}">d</figure-callout>}}_1+d_2+d_3)---\left(5\right)$

$H_{i,j}=\sqrt{L_{i,j}(L_{i,j}-d_1)(L_{i,j}-d_2)(L_{i,j}-d_3)}---\left(6\right)$

$R_{i,j}=\frac{H_{i,j}}{L_{i,j}}---\left(7\right)$

Step 3-3. combines radius information as the proper vector D of this point _i=(R _{I, i}, R _{I, i+1}..., R _{I, N}, R _{I, 1}..., R _{I, i-1})

Step 4. generates the segmented shape feature descriptor.

Step 4-1 specifies the width S of a local average, the proper vector D that then step 3 is obtained _iBe divided into

Section, the length of each section is all S, the configuration sampling point sequence number of each section is respectively: [1, S], [S+1,2S] ..., [N-S+1, N].

Step 4-2 calculates weighted mean value to the eigenwert in each section, wherein

The weight of each point in expression S, j=1,2 ..., M represents the segments of point.

${\mathrm{dS}}_i^j=\frac{1}{S}\underset{t=(j-1)S+1}{\overset{\mathrm{jS}}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i^t{d}_i^t---\left(2\right)$

Step 4-3 configuration sampling point x _iThe New Characteristics vector can be expressed as

All sampled point x on step 5. pair profile _i(i=1,2 ..., N) reuse the operation of step 4 and calculate its distance feature vector DS _i, consist of distance feature matrix D S (X)=(DS ₁, DS ₂..., DS _N).

Step 6. adopts the row maximal value to carry out normalization for every delegation of eigenmatrix DS (X), and the eigenwert after normalization is between [1,1].

${\mathrm{dS}}_i^j=\frac{{\mathrm{dS}}_i^j}{{\max }_{t=\mathrm{1,2},...N}\left\{\right|{\mathrm{dS}}_t^j\left|\right\}}---\left(3\right)$

Step 7. is calculated the similarity between shape.

Step 7-1X and Y represent respectively the configuration sampling point set of shape A and B.On profile, arbitrary sampled point can be expressed as: x _i∈ X (i=1,2 ..., N), y _j∈ Y (j=1,2 ..., L), wherein N and L are the sampled point number;

Step 7-2 utilizes L ₁The tolerance formula carries out proper vector relatively:

$c(x_i,y_j)=\underset{t=1}{\overset{M}{\mathrm{\&Sigma;}}}{w}_t|{\mathrm{dS}}_{x_i}^t-{\mathrm{dS}}_{y_j}^t|---\left(4\right)$

w _tThe weight that represents each local feature value, account form is

Because at p _iAnd q _jPoint on every side is more important than the point of distance, so will give larger weight.

Step 7-3 is for profile X={x _iAnd Y={y _j, by the calculating of step 7-2 formula, can obtain similarity matrix W:

$W=\left[\begin{array}{ccc}c(x_1,y_1)& ...& ...\\ ...& c(x_i,y_i)& ...\\ ...& ...& c(x_N,y_L)\end{array}\right]$

Step 7-4 seeks the corresponding relation g (x) between X and Y: X → Y, when point one by one at once

Minimum.

Similarity between step 7-5 shape A and B can be expressed as:

The present invention has translation, rotational invariance, and calculates simply, and dimension is lower, and is more stable for local deformation.Only need simple segmentation modeling and data to calculate, just can be described and matching operation the plane objective trait, greatly reduce operation time, increased its versatility.

Description of drawings

Fig. 1 is that method of the present invention carries out to certain shape the diagram that segmentation represents.

Fig. 2 (a) is the example of shape profile up-sampling point, p ₁, p ₂, p ₃Be three sampled points wherein.Fig. 2 (b)-Fig. 2 (d) is respectively the characteristic spectrum of three points, and transverse axis is the sequence of point, and the longitudinal axis is Euclidean distance tolerance.Fig. 2 (b) is a p ₁Characteristic spectrum, comprise positive negative in feature.Fig. 2 (c) is a p ₂Characteristic spectrum, all other point is all at p ₂The top of place's tangent line is so eigenwert is all positive number.Fig. 2 (d) is a p ₃Characteristic spectrum, all other point is all at p ₃The below of place's tangent line is so eigenwert is all negative.

Fig. 3 is the picture specification of triangle radial features descriptor.

Embodiment

For making purpose of the present invention, technical scheme and advantage clearer, below in conjunction with accompanying drawing and instantiation, the present invention is described in further details.These examples are only illustrative, and are not limitation of the present invention.

1. read in a width shape picture, as the figure of " heart " shape in Fig. 1, be designated as X.Use the canny edge detection algorithm to extract shape edges.Adopt afterwards the uniform sampling method, extract 100 sampled points, X={x on profile ₁, x ₂..., x ₁₀₀, as shown in Figure 1.

2. calculate its metric as an example of arbitrary sampled point example, as the sampled point of the middle recess in Fig. 1.Here take calculate Euclidean distance as metric as example, as shown in fig. 1, represent that this sampled point is to the range information of all other sampled points, in order to retrain the relation of this point and the shape overall situation.For the validity of Enhanced feature, added " positive negative direction " for Euclidean distance, namely other sample is in the left side of the tangent line of current point, and distance is for just; Distance is negative on the right side.Three kinds of possible situations have been showed as Fig. 2, for p ₁, there is sampled point simultaneously in its tangent line both sides, thus its proper vector have just have negative, as Fig. 2 (b); For p ₂, all sampled points all are positioned on the left of its tangent line, so its eigenwert is all positive number, as Fig. 2 (c); On the contrary, for p ₃, all sampled points all are positioned at its tangent line right side, so its all eigenwerts are all negative, as Fig. 2 (d).

3. can obtain the proper vector of each sampled point by calculating top calculating, as formula (1).

4. owing to there being a large amount of local deformations between shape, in order to overcome deformation effect, the present invention adopts local segmentation and average weighted method.At first whole profile is divided into 20 segments, every section has 5 sampled points.Afterwards the metric of 5 sampled points is weighted summation, as formula (2), the proper vector of each sampled point is reduced to 20 dimensions from 100 dimensions like this, has not only simplified the calculated amount of subsequent operation, and increased stability to local deformation, guaranteed simultaneously the validity of feature.

5. because European metric itself just has translation and invariable rotary feature, in order to increase the yardstick unchangeability of proper vector, the present invention has adopted capable maximum method for normalizing, and intersegmental feature is carried out normalization, as formula (3).

6. the proper vector with each sampled point of shape X connects together the eigenmatrix that has just consisted of whole shape.Use afterwards formula (4) to calculate similarity between any two sampled points of two shapes, generate similarity matrix W, utilize at last dynamic programming algorithm, find the point correspondence between two shapes, corresponding similarity value is accumulated together, just obtained the similarity size of two shapes.

7. the process of form fit is exactly that all the shape pictures in inquiry shape picture and database are calculated similarity one by one, then according to similarity size order rank results.

8. the example of triangle radial features descriptor and Euclidean distance are similar, just in the 2nd step difference to some extent during the calculated characteristics vector.The computation process of triangle radius as shown in Figure 3.At first utilize left and right consecutive point and arbitrary other sampled point of current sampling point to consist of a triangle, as shown in three solid lines.Then according to formula (5) (6) (7), derive the value of triangle radius.

Claims (2)

1. shape matching method based on metric, its feature comprises the following steps,

Step 1. pair original image utilizes the method for rim detection to extract profile;

Step 2. is carried out uniform sampling on profile, sampling number is N; N configuration sampling point is expressed as X={x ₁, x ₂..., x _N;

Step 3. is calculated arbitrary configuration sampling point x _i(i=1,2 ..., metric N);

Step 3-1 is with x _iBe starting point, scan counterclockwise other configuration sampling points, calculate x _iEuclidean distance (Euclidean Distance, ED) to other all configuration sampling points;

Step 3-2 is from x _iPoint is done the tangent line of profile, and in the tangent line left side, distance is being for just when other configuration sampling point, and distance is to bear on the tangent line right side;

The x that step 3-3 will calculate above _iPoint forms following proper vector D to the distance of other configuration sampling points _i

D _i=(d _i，i,d _i，i+1,...,d _i,N,d _i，1,...,d _i，i-1) （1）

Wherein, d _{I, j}Expression x _iTo x _jDistance, d _{I, i}Value is 0;

Step 4. generates the segmented shape feature descriptor;

Step 4-1 specifies the width S of a local average, the proper vector D that then step 3 is obtained _iBe divided into

Section, the length of each section is all S, the configuration sampling point sequence number of each section is respectively: [1, S], [S+1,2S] ..., [N-S+1, N];

Step 4-2 calculates weighted mean value to the eigenwert in each section, wherein The weight of each point in expression S, j=1,2 ..., M represents the segments of point;

${\mathrm{dS}}_i^j=\frac{1}{S}\underset{t=(j-1)S+1}{\overset{\mathrm{jS}}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i^t{d}_i^t---\left(2\right)$

Step 4-3 configuration sampling point x _iThe New Characteristics vector can be expressed as

All sampled point x on step 5. pair profile _i(i=1,2 ..., N) reuse the operation of step 4 and calculate its distance feature vector DS _i, consist of distance feature matrix D S (X)=(DS ₁, DS ₂..., DS _N);

Step 6. adopts the row maximal value to carry out normalization for every delegation of eigenmatrix DS (X), and the eigenwert after normalization is between [1,1];

${\mathrm{dS}}_i^j=\frac{{\mathrm{dS}}_i^j}{{\max }_{t=\mathrm{1,2}...N}\left\{\right|{\mathrm{dS}}_t^j\left|\right\}}---\left(3\right)$

Step 7. is calculated the similarity between shape;

Step 7-1X and Y represent respectively the configuration sampling point set of shape A and B.On profile, arbitrary sampled point can be expressed as: x _i∈ X (i=1,2 ..., N), y _j∈ Y (j=1,2 ..., L), wherein N and L are the sampled point number;

Step 7-2 utilizes L ₁The tolerance formula carries out proper vector relatively:

$c(x_i,y_j)=\underset{t=1}{\overset{M}{\mathrm{\&Sigma;}}}{w}_t|{\mathrm{dS}}_{x_i}^t-{\mathrm{dS}}_{y_j}^t|---\left(4\right)$

w _tThe weight that represents each local feature value, account form is

Step 7-3 is for profile X={x _iAnd Y={y _j, by the calculating of step 7-2 formula, can obtain similarity matrix W:

$W=\left[\begin{array}{ccc}c(x_1,y_1)& ...& ...\\ ...& c(x_i,y_j)& ...\\ ...& ...& c(x_N,y_L)\end{array}\right]$

Step 7-4 seeks the corresponding relation g (x) between X and Y: X → Y, when point one by one at once

Minimum;

Similarity between step 7-5 shape A and B is expressed as:

2. shape matching method according to claim 1, is characterized in that, can with the triangle radius as metric, namely use following step replacement step 3 for the calculating to metric in step 3:

Step 3-1. is for the sampled point x on profile _i, some x _i-1And x _i+1Respectively its left and right consecutive point, other arbitrary point x on profile _jCan both and x _i-1And x _i+1Consist of a triangle; If d ₁Expression x _i-1And x _jBetween Euclidean distance, d ₂Expression x _i+1And x _jBetween distance, d ₃Expression x _i-1And x _i+1Between distance;

Step 3-2. calculates leg-of-mutton semi-perimeter L with following formula _{I, j}, area H _i,jAnd radius R _{I, j}:

$L_{i,j}=\frac{1}{2}(d_1+d_2+d_3)---\left(5\right)$

$H_{i,j}=\sqrt{L_{i,j}(L_{i,j}-d_1)(L_{i,j}-d_2)(L_{i,j}-d_3)}---\left(6\right)$

$R_{i,j}=\frac{H_{i,j}}{L_{i,j}}---\left(7\right)$

Step 3-3. combines radius information as the proper vector D of this point _i=(R _{I, i}, R _{I, i+1}..., R _{I, N}, R _{I, 1}..., R _{I, i-1}).

Images