CN112633304B - Robust fuzzy image matching method - Google Patents

Abstract

The invention relates to a robust fuzzy image matching method. The method comprises the following steps: two images with different blur degree are input first. Secondly, a group of scale-invariant feature transform (SIFT) points are extracted, and three scale-invariant concentric circle regions are applied to generate descriptors in order to further improve the specificity of SIFT descriptors. Third, to reduce the high-dimensional complexity of SIFT descriptors, local retention projection LPP techniques are employed to reduce the size of the descriptors. And finally, obtaining matching characteristic points by utilizing Euclidean distance similarity measurement. The method not only can reduce the data quantity, but also can improve the matching speed and the matching precision, and can be suitable for other image matching methods.

Description

Robust fuzzy image matching method

Technical Field

The invention relates to the technical field of computer vision, in particular to a robust fuzzy image matching method.

Background

Image matching is a special field of image processing, corresponding geometric relations among images are determined by extracting consistent feature points among different images of the same scene through image matching, a matched image is obtained, the image scene can be described more accurately than a single image, generally, the image matching can be carried out by adopting a method based on local feature extraction and matching, the method for extracting and matching the local features mainly considers the scale and rotation invariance of an input image, larger calculated data quantity and instantaneity are not considered, and corresponding matching point pairs can not be obtained effectively and accurately for image matching in a fuzzy scene.

Disclosure of Invention

Aiming at the defects in the prior art, the invention provides a robust fuzzy image matching method, which utilizes a three-scale center unchanged circular area and an LPP technology to reduce the dimension of a descriptor, greatly improves the operation efficiency while enhancing the distinguishing property of characteristic points, and greatly improves the correct matching rate and the robustness.

The technical scheme adopted by the invention for achieving the purpose is as follows: a robust blurred image matching method comprising the steps of:

s1: inputting two original images with different blurring degrees;

s2: extracting feature points on two original images by using a Scale Invariant Feature Transform (SIFT) algorithm;

s3: respectively establishing three central circular areas with unchanged scales around the characteristic points on the two original images, and describing the characteristic points to form respective characteristic point descriptors of the two original images;

s4: the local projection mapping LPP method is adopted to reduce the dimension of the feature point descriptor for improving the arithmetic operation efficiency;

s5: and matching the feature point descriptors of the two original images after the dimension reduction, and selecting accurate matching point pairs from the two images.

The feature points are described in step S3 as directivity information specifying descriptors.

The directivity information of the specified descriptor includes:

describing each characteristic point by using 16 seed points 4×4, dividing a gradient histogram of an area where each seed point is located into 8 direction intervals between 0 ° and 360 °, and carrying out weighting operation on the gradient histogram by using a Gaussian window to generate 128-dimensional characteristic vectors;

defining a feature point descriptor LSIFT described by three scale-invariant central regions is expressed as:

PD＝α ₁ L ₁ +α ₂ L ₂ +α ₃ L ₃

wherein L is _i (i=1, 2, 3) is a 128-dimensional SIFT descriptor, PD is a weighted 128-dimensional descriptor, α ₁ ,α ₂ ,α ₃ Is a preset weighting coefficient.

The step S4 of reducing feature point descriptor dimensions by applying the local projection mapping LPP method includes:

a. defining a feature point descriptor LSIFT described by three scale-invariant central regions as (x ₁ ,x ₂ ,…x _m ),x _i A feature point descriptor LSIFT representing one of the images; y is _i ＝w ^T x _i Representing a one-dimensional description of the transformation vector w, defining a similarity matrix S (S _ij ＝s _ji )：

b. Selecting an appropriate projection is to minimize the solution of the objective function f:

wherein D is a diagonal matrix D _ii ＝∑ _j S _ij l=d-S is a laplace matrix. There are the following constraints:

Y ^T DY＝w ^T XDX ^T w＝1

c. the problem of minimizing the solution of the objective function f can be reduced to:

w ^T XDX ^T w＝1

d. can be converted into a generalized eigenvalue problem:

XLX ^T w＝λXDX ^T w

wherein XLX ^T ,XDX ^T Are symmetrical and semi-positive definite matrices;

e. let W be the column vector of the generalized eigenvalue λ, projection matrix W _LPP ＝(w ₀ ,w ₁ ,…w _l-1 ) Each vector w of PD _i (i=0, 1, …, l-1) all have 128 dimensions, then the projection matrix reduces the 128-dimensional descriptor vector to l-dimensions, so the 128-dimensional descriptor is translated into:

TPD＝PD·W _LPP

where TPD is a local descriptor of dimension l < 128.

The descriptor matching in the step S5 includes:

computing two descriptors TPD _i ,TPD _j The Euclidean distance between the two is obtained by adopting a nearest neighbor and next nearest neighbor algorithm to obtain an accurate matching point pair:

D _{nearest neighbor} /D _{Secondary nearest neighbor} ＜T

Wherein D is _{Nearest neighbor} And D _{Secondary nearest neighbor} The nearest neighbor distance and the next nearest neighbor distance when the current pixel point is taken as an origin are respectively represented, T represents a threshold value used for matching, and the current pixel point is a matching point pair.

D described in step S5 _{Nearest neighbor} And D _{Secondary nearest neighbor} The calculation is carried out according to the following formula:

wherein, TPD _i Descriptor representing arbitrary feature point i after dimension reduction, TPD _j Descriptor, TPD, representing feature point j after dimension reduction _i,m Mth dimension vector representing feature point i descriptor, TPD _j,m The m-th dimension vector representing the descriptor of the feature point j, i represents the dimension after dimension reduction.

The invention has the following beneficial effects and advantages:

1. the robust fuzzy image matching method describes the characteristic points by means of the three central circular areas with unchanged scales, enhances the distinguishing property of the characteristic descriptors and improves the correct matching rate.

2. The robust fuzzy image matching method reduces the dimension of the feature point descriptor by utilizing the local projection mapping technology, improves the matching efficiency of the image on the basis of ensuring the correct matching rate, and has stronger instantaneity.

Drawings

FIG. 1 is a flow chart of the method of the present invention;

FIG. 2 is a blurred image of a structural scene of the method of the present invention;

FIG. 3 is a graph showing the effect of different parameters of a structural type image of the method of the present invention on blurred image performance;

FIG. 4 is a blurred image of a texture scene of the method of the present invention;

FIG. 5 is a graph showing the effect of different parameters of a blurred image of a texture scene on the performance of the blurred image in accordance with the method of the present invention;

Detailed Description

The present invention will be described in further detail with reference to the accompanying drawings.

Referring to fig. 1, the specific steps of a robust blurred image matching method of the present invention are as follows:

step 1: inputting two original images with different blurring degrees;

step 2: extracting feature points on two original images by using a Scale Invariant Feature Transform (SIFT) algorithm;

step 3: respectively establishing three central circular areas with unchanged scales around the characteristic points on the two original images, and describing the characteristic points to form respective characteristic point descriptors of the two original images;

describing each characteristic point by using 16 seed points 4×4, dividing a gradient histogram of an area where each seed point is located into 8 direction intervals between 0 ° and 360 °, and carrying out weighting operation on the gradient histogram by using a Gaussian window to generate 128-dimensional characteristic vectors;

defining a feature point descriptor LSIFT described by three scale-invariant central regions is expressed as:

PD＝α ₁ L ₁ +α ₂ L ₂ +α ₃ L ₃

wherein L is _i (i=1, 2, 3) is a 128-dimensional SIFT descriptor, PD is a weighted 128-dimensional descriptor, α ₁ ,α ₂ ,α ₃ Is a preset weighting coefficient.

Step 4: in order to improve the algorithm operation efficiency, the dimension of the feature point descriptor is reduced by applying a local projection mapping (LPP) method;

defining a feature point descriptor LSIFT described by three scale-invariant central regions as (x ₁ ,x ₂ ,…x _m ),x _i A LSIFT representing one of the images; y is _i ＝w ^T x _i Representing a one-dimensional description of the transformation vector w, defining a similarity matrix S (S _ij ＝s _ji )：

b. Selecting an appropriate projection is to minimize the solution of the objective function f:

wherein D is a diagonal matrix D _ii ＝∑ _j S _ij l=d-S is a laplace matrix. There are the following constraints:

Y ^T DY＝w ^T XDX ^T w＝1

c. the problem of minimizing the solution of the objective function f can be reduced to:

w ^T XDX ^T w＝1

d. can be converted into a generalized eigenvalue problem:

XLX ^T w＝λXDX ^T w

wherein XLX ^T ,XDX ^T Are symmetrical and semi-positive definite matrices;

e. let W be the column vector of the generalized eigenvalue λ, projection matrix W _LPP ＝(w ₀ ,w ₁ ,…w _l-1 ) Each w _i (i=0, 1, …, l-1) all have 128 dimensions, then the projection matrix reduces the 128-dimensional descriptor vector to l-dimensions, so the 128-dimensional descriptor is translated into:

TPD＝PD·W _LPP

where TPD is a local descriptor l < 128 in one dimension.

And 5, matching the feature point descriptors of the two original images after the dimension reduction, and selecting accurate matching point pairs from the two images.

Computing two descriptors TPD _i ,TPD _j The Euclidean distance between the two is obtained by adopting a nearest neighbor and next nearest neighbor algorithm to obtain an accurate matching point pair:

D _{nearest neighbor} /D _{Secondary nearest neighbor} ＜T

Wherein D is _{Nearest neighbor} And D _{Secondary nearest neighbor} The nearest neighbor distance and the next nearest neighbor distance when the current pixel point is taken as an origin are respectively represented, and T represents a threshold value used for matching. Wherein:

wherein, TPD _i Descriptor representing arbitrary feature point i after dimension reduction, TPD _j Descriptor indicating feature point j after dimension reduction, j not including i, TPD _i,m Mth dimension vector representing feature point i descriptor, TPD _j,m The m-th dimension vector representing the descriptor of the feature point j, l represents the dimension after dimension reduction.

The effects of the present invention are further described below with reference to the accompanying drawings.

In order to verify the validity and the correctness of the invention, two groups of fuzzy images of a structural scene and a texture scene are adopted to carry out a matching simulation experiment. All simulation experiments were implemented under Windows XP operating system using Visual Studio2010 software.

Simulation example 1:

fig. 2 shows blurred images of six structural scenes obtained under different blur degrees, wherein the adopted image size is 800×600, wherein (a) images are reference images, and other (b) - (f) images are images to be matched respectively. Fig. 3 (a) shows the correct matching number of the structural image, the horizontal axis shows the degree of blurring, the vertical axis shows the number of correct matching points, fig. 3 (b) shows the correct matching rate of the structural image, the horizontal axis shows the degree of blurring, and the vertical axis shows the correct matching rate; as can be seen from fig. 3 (a) and fig. 3 (b), the number of correct matching points obtained by the method of the present invention under all fuzzy variation conditions is significantly higher than that of the SIFT method.

Simulation example 2:

fig. 4 shows blurred images of six texture scenes obtained under different blur degrees, wherein the adopted image size is 800×600, wherein (a) images are reference images, and other (b) - (f) images are images to be matched respectively. Fig. 5 (a) shows the correct matching number of the texture image, the horizontal axis shows the degree of blurring, the vertical axis shows the number of correct matching points, fig. 5 (b) shows the correct matching rate of the texture image, the horizontal axis shows the degree of blurring, and the vertical axis shows the correct matching rate; it can be seen from fig. 5 (a) and fig. 5 (b) that the number of correct matching points obtained by the method of the present invention under all fuzzy variation conditions is significantly higher than that of the SIFT method.

The invention can accurately match the image with fuzzy change, not only can obtain higher matching point pairs, but also has higher correct matching rate.

Claims (5)

1. A robust blurred image matching method, comprising the steps of:

s1: inputting two original images with different blurring degrees;

s2: extracting feature points on two original images by using a Scale Invariant Feature Transform (SIFT) algorithm;

s3: respectively establishing three central circular areas with unchanged scales around the characteristic points on the two original images, and describing the characteristic points to form respective characteristic point descriptors of the two original images;

s4: the local projection mapping LPP method is adopted to reduce the dimension of the feature point descriptor for improving the arithmetic operation efficiency; comprising the following steps:

a. defining a feature point descriptor LSIFT described by three scale-invariant central regions as (x ₁ ,x ₂ ,…x _m ),x _i A feature point descriptor LSIFT representing one of the images; y is _i ＝w ^T x _i Representing a one-dimensional description of the transformation vector w, defining a similarity matrix S (S _ij ＝s _ji )：

b. Selecting an appropriate projection is to minimize the solution of the objective function f:

wherein D is a diagonal matrix D _ii ＝∑ _j S _ij l=d-S is a laplace matrix; there are the following constraints:

Y ^T DY＝w ^T XDX ^T w＝1

c. the problem of minimizing the solution of the objective function f is simplified as:

w ^T XDX ^T w＝1

d. conversion to generalized eigenvalue problem:

XLX ^T w＝λXDX ^T w

wherein XLX ^T ,XDX ^T Are symmetrical and semi-positive definite matrices;

e. let W be the column vector of the generalized eigenvalue λ, projection matrix W _LPP ＝(w ₀ ,w ₁ ,…w _l-1 ) Each vector w of PD _i (i=0, 1, …, l-1) all have 128 dimensions, then the projection matrix reduces the 128-dimensional descriptor vector to l-dimensions, so the 128-dimensional descriptor is translated into:

TPD＝PD·W _LPP

wherein TPD is a local descriptor of dimension l < 128;

s5: and matching the feature point descriptors of the two original images after the dimension reduction, and selecting accurate matching point pairs from the two images.

2. A robust blurred image matching method as claimed in claim 1 wherein said describing feature points in step S3 is specifying directional information of descriptors.

3. A robust blurred image matching method as claimed in claim 2 wherein said directional information specifying descriptors includes:

describing each characteristic point by using 16 seed points 4×4, dividing a gradient histogram of an area where each seed point is located into 8 direction intervals between 0 ° and 360 °, and carrying out weighting operation on the gradient histogram by using a Gaussian window to generate 128-dimensional characteristic vectors;

defining a feature point descriptor LSIFT described by three scale-invariant central regions is expressed as:

PD＝α ₁ L ₁ +α ₂ L ₂ +α ₃ L ₃

wherein L is _i (i=1, 2, 3) is a 128-dimensional SIFT descriptor, PD is a weighted 128-dimensional descriptor, α ₁ ,α ₂ ,α ₃ Is a preset weighting coefficient.

4. A robust blurred image matching method as claimed in claim 1 wherein said descriptor matching in step S5 comprises:

computing two descriptors TPD _i ,TPD _j The Euclidean distance between the two is obtained by adopting a nearest neighbor and next nearest neighbor algorithm to obtain an accurate matching point pair:

D _{nearest neighbor} /D _{Secondary nearest neighbor} ＜T

Wherein D is _{Nearest neighbor} And D _{Secondary nearest neighbor} The nearest neighbor distance and the next nearest neighbor distance when the current pixel point is taken as an origin are respectively represented, T represents a threshold value used for matching, and the current pixel point is a matching point pair.

5. The method of claim 4, wherein said step S5 is characterized by said step D _{Nearest neighbor} And D _{Secondary nearest neighbor} The calculation is carried out according to the following formula:

wherein, TPD _i Descriptor representing arbitrary feature point i after dimension reduction, TPD _j Descriptor, TPD, representing feature point j after dimension reduction _i,m Mth dimension vector representing feature point i descriptor, TPD _j,m The m-th dimension vector representing the descriptor of the feature point j, i represents the dimension after dimension reduction.