CN104239531A - Accurate comparison method based on local visual features - Google Patents

Abstract

The invention relates to an accurate comparison method based on local visual features, which comprises the following steps of (1) acquiring a Query vehicle image and multiple database images to be compared; (2) carrying out key point extraction on all images, and expressing each image to be a key point set; (3) generating a feature tree of the Query vehicle image according to the key point set; (4) matching the feature tree with the database images in the step (1), enabling a defined objective function to be the smallest, and outputting the similarity of the Query vehicle image and each database image. Compared with the prior art, the accurate comparison method based on the local visual features has the advantages of high accuracy, good robustness and the like.

Description

A kind of precise alignment method based on local visual feature

Technical field

The present invention relates to a kind of image comparison method, especially relate to a kind of precise alignment method based on local visual feature.

Background technology

In recent years, intelligent transportation system development fast, along with the development of computer vision and mode identification technology, for the more effective application of intelligent transportation system provides opportunity.Computer vision utilizes computing machine to simulate the visual performance of people, information extraction from the image of objective things, carries out process and understood, and finally detects for reality, measures and control.

Existing image alignments, for the consideration of speed aspect, all abandons the information of Shape aspect.Often use the mode of RANSAC to add Shape information in current image detecting system, but when it uses, the texture information (proper vector) of unique point and Shape are divided into two stages considerations, therefore performance improves not obvious.

Summary of the invention

Object of the present invention be exactly in order to overcome above-mentioned prior art exist defect and a kind of precise alignment method based on local visual feature that degree of accuracy is high, robustness is good is provided.

Object of the present invention can be achieved through the following technical solutions:

Based on a precise alignment method for local visual feature, comprise the following steps:

1) Query vehicle image to be compared and multiple database images are obtained;

2) key point extraction being carried out to all images, being all expressed as set of keypoints by often opening image;

3) generate the characteristics tree of Query vehicle image according to set of keypoints, be specially:

31) in the set of keypoints that Query vehicle image is corresponding, extract 1 point, itself and immediate 2 points are formed a triangle;

32) remove 3 points corresponding to described triangle in set of keypoints, and described leg-of-mutton mid point is added in set of keypoints as key point;

33) repeat step 31), 32), until remaining 1 point of set of keypoints;

34) according to step 31)-33) institute have some morphogenesis characters tree;

4) by described characteristics tree and step 1) in database images mate, make the objective function that defines minimum, export Query vehicle image and the similarity of often opening database images, similarity and objective function are in the value of characteristics tree root node.

Described step 2), key point extraction is carried out to image and is specially:

21) car plate detection and brand recognition process is carried out to often opening vehicle image;

22) extract vehicle image according to the car plate position detected, and be normalized;

23) adopt various features extracting method to carry out feature point extraction to the vehicle image of often opening extracted, non-maxima suppression process is carried out to the unique point that distinct methods obtains, only retains a key point in same area.

The element number of the initial key point set of described Query vehicle image is odd number.

In described characteristics tree, each key point represents with circle, and the mid point after 3 points form triangle represents with triangle.

Described objective function is:

For circle, objective function is the L2 distance of proper vector, as follows:

$O_p=\underset{l}{\mathrm{\Σ}}{(f_{\mathrm{il}}-f_{\mathrm{jl}})}^2$

In formula, p is present node, f _ifor Query vehicle image is in the proper vector of this point, f _jdatabase images is at the proper vector f of this point _il, f _il, f _jlbe respectively vector f _i, f _jin l element;

For triangle, objective function is made up of 2 parts, and comprise the summation of the Score forming leg-of-mutton 3 child nodes and the similarity between 3 child nodes formation triangles and Query vehicle image formation triangle, concrete formula is:

$O_p=\underset{\mathrm{ch}=1}{\overset{3}{\mathrm{\Σ}}}{O}_{\mathrm{ch}}+\underset{l=1}{\overset{R}{\mathrm{\Σ}}}(g_{\mathrm{il}}-g_{\mathrm{jl}})$

In formula, be the summation of the Score of 3 child nodes, g _ifor the leg-of-mutton shape eigenvectors that Query vehicle image is formed, g _jfor the leg-of-mutton shape eigenvectors that 3 child nodes of database images are formed, R is g _i, g _jlength.

Described step 4) in, when mating, the concrete disposal route of dissimilar node is as follows:

Circular: contrast is likely gathered, retain K the most similar possibility;

Triangle: linear combination, each child nodes have K may, total K*K*K kind may, for each may calculating target function value, finally retain K the possibility that Score is maximum;

Root node: get the highest may conduct the exporting of Score.

Described step 4) in, adopt mode from top to bottom when mating.

Compared with prior art, the present invention has the following advantages:

1, performance is good, and on the database of 1,000,000 ranks, accuracy can close to 9 one-tenth;

2, strong robustness, can use under different scene.

Accompanying drawing explanation

Fig. 1 is schematic flow sheet of the present invention;

Fig. 2 is the schematic diagram of characteristics tree of the present invention.

Embodiment

Below in conjunction with the drawings and specific embodiments, the present invention is described in detail.The present embodiment is implemented premised on technical solution of the present invention, give detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to following embodiment.

As shown in Figure 1, a kind of precise alignment method based on local visual feature, comprises the following steps:

In step s1, obtain Query vehicle image to be compared and multiple database images (database image), as input.

In step s2, key point extraction is carried out to all images, being all expressed as set of keypoints by often opening image, being specially:

21) car plate detection and brand recognition process is carried out to often opening vehicle image;

22) extract vehicle image according to the car plate position detected, and be normalized;

23) adopt various features extracting method to carry out feature point extraction to the vehicle image of often opening extracted, non-maxima suppression process is carried out to the unique point that distinct methods obtains, only retains a key point in same area.

In step s3, generate the characteristics tree of Query vehicle image according to set of keypoints, be specially:

31) in the set of keypoints that Query vehicle image is corresponding, extract 1 point, itself and immediate 2 points are formed a triangle;

32) remove 3 points corresponding to described triangle in set of keypoints, and described leg-of-mutton mid point is added in set of keypoints as key point;

33) repeat step 31), 32), until remaining 1 point of set of keypoints, the element number needing the initial key point set ensureing Query vehicle image is odd number, often takes turns iteration and reduces by 2 points, last certain only residue 1 point;

34) according to step 31)-33) institute have some morphogenesis characters tree, as shown in Figure 2, in characteristics tree, each key point represents with circle, and the mid point after 3 points form triangles represents with triangle.

In step s4, described characteristics tree mated with the database images in step s1, make the objective function that defines minimum, export Query vehicle image and the similarity of often opening database images, similarity and objective function are in the value of characteristics tree root node.

For often opening database image Ai, the objective function finding a coupling to make to define is minimum.Wherein, mating definition is: find out the position on this figure Ai for each circle in exemplary plot.Objective function is defined as follows:

For circle, objective function is the L2 distance of proper vector, as follows:

$O_p=\underset{l}{\mathrm{\Σ}}{(f_{\mathrm{il}}-f_{\mathrm{jl}})}^2$

In formula, p is present node, f _ifor Query vehicle image is in the proper vector of this point, f _jdatabase images is at the proper vector f of this point _il, f _il, f _jlbe respectively vector f _i, f _jin l element.

For triangle: objective function is made up of 2 parts, comprise the summation of the Score forming leg-of-mutton 3 child nodes and the similarity between 3 child nodes formation triangles and Query image construction triangle.The explanation of formula 3 in document " L.Zhu; Y.Chen; C.Lin; A.L.Yuille.Rapid Inference on a novel AND/OR graph:Detection; Segmentation and Parsing of Articulated Deformable Objects in Cluttered Backgrounds.Advances in Neural Information Processing Systems 20.NIPS.December2007 " is shown in its definition, is specially:

$O_p=\underset{\mathrm{ch}=1}{\overset{3}{\mathrm{\Σ}}}{O}_{\mathrm{ch}}+\underset{l=1}{\overset{R}{\mathrm{\Σ}}}(g_{\mathrm{il}}-g_{\mathrm{jl}})$

In formula, be the summation of the Score of 3 child nodes, g _ifor the leg-of-mutton shape eigenvectors that Query vehicle image is formed, g _jfor the leg-of-mutton shape eigenvectors that 3 child nodes of database images are formed, R is g _i, g _jlength.

Adopt mode from top to bottom when mating, use the method for dynamic programming, first process child nodes, reprocessing father node, the concrete disposal route of dissimilar node is as follows:

Circular: contrast is likely gathered, retain K the most similar possibility;

Triangle: linear combination, each child nodes have K may, total K*K*K kind may, for each may calculating target function value, finally retain K the possibility that Score is maximum;

Root node: get the highest may conduct the exporting of Score.

Claims (7)

1., based on a precise alignment method for local visual feature, it is characterized in that, comprise the following steps:

1) Query vehicle image to be compared and multiple database images are obtained;

2) key point extraction being carried out to all images, being all expressed as set of keypoints by often opening image;

3) generate the characteristics tree of Query vehicle image according to set of keypoints, be specially:

31) in the set of keypoints that Query vehicle image is corresponding, extract 1 point, itself and immediate 2 points are formed a triangle;

32) remove 3 points corresponding to described triangle in set of keypoints, and described leg-of-mutton mid point is added in set of keypoints as key point;

33) repeat step 31), 32), until remaining 1 point of set of keypoints;

34) according to step 31)-33) institute have some morphogenesis characters tree;

4) by described characteristics tree and step 1) in database images mate, make the objective function that defines minimum, export Query vehicle image and the similarity of often opening database images, similarity and objective function are in the value of characteristics tree root node.

2. a kind of precise alignment method based on local visual feature according to claim 1, is characterized in that, described step 2), key point extraction is carried out to image and is specially:

21) car plate detection and brand recognition process is carried out to often opening vehicle image;

22) extract vehicle image according to the car plate position detected, and be normalized;

23) adopt various features extracting method to carry out feature point extraction to the vehicle image of often opening extracted, non-maxima suppression process is carried out to the unique point that distinct methods obtains, only retains a key point in same area.

3. a kind of precise alignment method based on local visual feature according to claim 1, is characterized in that, the element number of the initial key point set of described Query vehicle image is odd number.

4. a kind of precise alignment method based on local visual feature according to claim 1, it is characterized in that, in described characteristics tree, each key point represents with circle, and the mid point after 3 points form triangle represents with triangle.

5. a kind of precise alignment method based on local visual feature according to claim 4, it is characterized in that, described objective function is:

For circle, objective function is the L2 distance of proper vector, as follows:

$O_p=\underset{l}{\mathrm{\Σ}}{(f_{\mathrm{il}}-f_{\mathrm{jl}})}^2$

In formula, p is present node, f _ifor Query vehicle image is in the proper vector of this point, f _jdatabase images is at the proper vector f of this point _il, f _il, f _jlbe respectively vector f _i, f _jin l element;

For triangle, objective function is made up of 2 parts, and comprise the summation of the Score forming leg-of-mutton 3 child nodes and the similarity between 3 child nodes formation triangles and Query vehicle image formation triangle, concrete formula is:

$O_p=\underset{\mathrm{ch}=1}{\overset{3}{\mathrm{\Σ}}}{O}_{\mathrm{ch}}+\underset{l=1}{\overset{R}{\mathrm{\Σ}}}(g_{\mathrm{il}}-g_{\mathrm{jl}})$

In formula, be the summation of the Score of 3 child nodes, g _ifor the leg-of-mutton shape eigenvectors that Query vehicle image is formed, g _jfor the leg-of-mutton shape eigenvectors that 3 child nodes of database images are formed, R is g _i, g _jlength.

6. a kind of precise alignment method based on local visual feature according to claim 4, is characterized in that, described step 4) in, when mating, the concrete disposal route of dissimilar node is as follows:

Circular: contrast is likely gathered, and retains the most similar, K that namely objective function is minimum may;

Triangle: linear combination, each child nodes have K may, total K*K*K kind may, for each may calculating target function value, finally retain K the possibility that Score is maximum;

Root node: get the highest may conduct the exporting of Score.

7. a kind of precise alignment method based on local visual feature according to claim 1, is characterized in that, described step 4) in, adopt mode from top to bottom when mating.