CN104700400A - High-precision image matching method based on feature point training - Google Patents

Abstract

The invention relates to a high-precision image matching method based on feature point training, which performs feature point matching on images to be registered under different angles and standard images before image registration, and extracts features that can be accurately matched under different angles The point is used as the reference feature point in the actual detection process. The feature points selected by the invention are not only beneficial to search for feature points with higher positioning accuracy as the calculation basis of the transformation model, thereby improving the positioning accuracy; but also can further reduce the calculation amount of feature point matching operations and speed up the operation speed of the algorithm.

Description

A high-precision image matching method based on feature point training

technical field

The invention relates to a feature point matching method, in particular to a high-precision image matching method based on feature point training.

Background technique

Image registration is a basic problem in the field of machine vision, and it has always been a hot and difficult point of research. Image registration refers to finding correspondences between two or more images of the same scene from different times, different viewpoints, or different sensors. The methods related to image registration are mainly divided into: methods based on grayscale registration and methods based on feature registration.

The method based on grayscale registration, also known as correlation registration method, uses spatial two-dimensional sliding templates for image registration. The difference between different algorithms is mainly reflected in the selection of relevant criteria. The feature point-based registration method first extracts features in the original image, and then establishes the registration relationship of feature points between the two images, which has high robustness, such as the SURF algorithm. At present, image registration based on SURF algorithm has been widely used, such as in medical image registration, remote sensing image registration, etc. The SURF algorithm first extracts the feature points in the image for feature point matching, and uses the matched feature point pairs to calculate the parameters of the image registration transformation model, so that the image to be registered can be accurately transformed to the standard image position to achieve the purpose of registration. The matching of feature points is the key to image registration, and the accuracy of matching directly affects the accuracy of subsequent registration. In the algorithm, according to the feature point neighborhood information contained in the SURF feature point descriptor, the K nearest neighbor method can be used to find out the two potential best matching points for each feature point, and the distance between the best value of the matching point and the second best The value ratio is optimized to find the best matching point pair.

However, in the actual image registration process, the image to be registered has changes at different angles relative to the standard image. Although the SURF feature point detection algorithm has rotation-invariant feature points, as the image changes at different angles, the extracted The positioning accuracy of feature points still has certain differences. In order to improve the accuracy of image registration, how to extract feature points that are not affected by image angle changes becomes the key to solving the problem.

Contents of the invention

The purpose of the present invention is to provide a high-precision image matching method based on feature point training, which selects feature points that are not affected by changes in image rotation angles, thereby improving the accuracy of image registration.

The present invention solves its technical problem and realizes by taking the following technical solutions:

A high-precision image matching method based on feature point training, which is characterized in that: before image registration, feature point matching is performed on images to be registered under different angles and standard images, and feature points with high positioning accuracy are extracted as actual The reference feature point in the detection process.

Moreover, the specific selection step of the reference feature point is: use the SURF feature point extraction algorithm to extract feature points in two images with different rotation angles, and respectively match the feature points in the standard image to obtain the matched feature points in the standard image. Feature point set, record the X and Y coordinates of the feature point; analyze the obtained two sets of feature point sets with successful matching, and select the feature point with exactly the same X and Y coordinates in the two feature point sets as the reference feature point.

Also, the rotation angle is 2° to 10°.

Moreover, the image registration adopts the SURF feature point detection algorithm, and the reference feature point is used as the calculation image registration transformation model parameter of the SURF algorithm image registration, so that the image to be registered is accurately transformed to the standard image position.

Advantage and positive effect of the present invention are:

1. By verifying the matching performance of feature points, the selected feature points can be accurately matched under different rotation conditions, and their coordinates reach the accuracy of three decimal places. These feature points have a strong ability to describe image features , with high positioning accuracy, and can better adapt to feature point matching under different rotation angles.

2. The feature points selected by the present invention are beneficial to search for feature points with higher positioning accuracy as the calculation basis of the transformation model, thereby improving the positioning accuracy; at the same time, it can further reduce the calculation amount of feature point matching operations and speed up the operation speed of the algorithm .

Description of drawings

Figure 1 shows the matching feature points in the counterclockwise 2-degree standard image;

Figure 2 is the matching feature points in the clockwise 2-degree standard image;

Figure 3 is a distribution diagram of feature points for matching in a standard image under different rotation angles;

Fig. 4 is the coordinate difference distribution result calculated by coincident feature points and non-overlapping feature points;

Figure 5 is a 9×9 box filter template;

Fig. 6 is the selection of the main direction of the feature point;

Figure 7. Composition of feature descriptors.

Detailed ways

The present invention will be further described in detail below in conjunction with the accompanying drawings and through specific embodiments. The following embodiments are only descriptive, not restrictive, and cannot limit the protection scope of the present invention.

A high-precision image matching method based on feature point training, which studies the degree to which the positioning accuracy of feature points is affected by different transformations. In this embodiment, the image to be tested that is rotated 2 degrees counterclockwise and the image to be tested that is rotated 2 degrees clockwise are selected for testing. experiment.

First, use the SURF feature point extraction algorithm to extract feature points in two images with different rotation angles, and match them with the feature points in the standard image to obtain the matched feature point set in the standard image, as shown in Figure 1 and Figure 2 Points of different shapes mark the matching feature points in the standard image under different rotation angles.

Subsequently, after further analyzing the two sets of feature point sets that were successfully matched, it was found that there were feature points with exactly the same X and Y coordinates in the two feature point sets (the two kinds of marks shown in Figure 3 completely overlapped points), X The feature point that is completely consistent with the Y coordinate is the feature point that can be accurately matched under different rotation angles.

Table 1 shows the coordinates of the partially coincident points.

Table 1 Coordinates of coincident points

serial number X coordinate y-coordinate 1 143.113 119.743 2 216.9 155.781 3 225.787 327.051 4 243.284 410.629 5 258.155 55.093 6 344.04 297.874 7 381.819 415.828 8 509.629 262.798 9 513.646 205.083 10 518.92 262.092

It can be seen from the experimental results that there are large differences in the set of feature points that are successfully matched under different rotation conditions. At the same time, we noticed that there are partial intersections in the above-mentioned feature point pair sets, that is, there are some feature points in the above-mentioned feature point sets that can be accurately matched under different rotation conditions, and their coordinates reach the accuracy of three decimal places. This experimental result shows that these feature points have a strong ability to describe image features, high positioning accuracy, and can better adapt to feature point matching under different rotation angles.

In order to further verify the image feature expression ability of this kind of feature points, the rotation adaptability experiment of feature points is carried out. Here, extract the coincident feature points and non-overlapping feature points shown in Figure 3 (both are feature points in the standard image) as feature point matching reference points, and the feature points extracted from the image to be tested rotated 10 degrees clockwise Perform feature point matching, and process the successfully matched feature points as follows:

According to the image rotation theory, calculate the theoretical coordinate values (obtained by theoretical calculation) of the matching feature points in the two sets of matching reference points after the image is rotated by 10 degrees;

Calculate the difference between the coordinate values of the feature points calculated in (1) and the coordinate values of the feature points that have been matched in the image to be tested rotated 10 degrees in the X and Y directions, so as to obtain the positioning deviation of the above feature points, The result is shown in Figure 4.

It can be seen from Figure 4 that the difference between the coordinate values of the coincident feature points extracted in the standard image and the theoretical coordinate values calculated in practice is relatively small, and relatively concentrated near the origin, indicating that its positioning accuracy is high. Correspondingly, the positioning deviation of non-coinciding feature points is relatively large. This result indicates that there is a certain difference in the localization accuracy of the feature points.

For this reason, before image registration, the image to be registered under different angles can be matched with the standard image for feature points, and the feature points with high positioning accuracy can be extracted as the reference feature points in the actual detection process. Through the above method, it is not only beneficial to search for feature points with higher positioning accuracy as the calculation basis of the transformation model, thereby improving the positioning accuracy; at the same time, it can further reduce the calculation amount of feature point matching operations and speed up the operation speed of the algorithm.

In order to illustrate that before image registration, the feature points in the standard image that can match the images to be registered under different angles can be extracted as the reference feature points in the actual detection process, and further analysis is carried out. By selecting two training images to be registered at different angles to match the feature points with the standard image, extract the feature points in the standard image that can match the two images to be registered, and use a test image from a different angle to perform Test to verify the validity of the extracted feature points. As shown in table 2.

Table 2 Analysis of feature point matching accuracy

It can be seen from Table 2 that the average distance between coincident points and theoretical coordinates is smaller than that of non-coincidence points, and is only about half of the latter. This further illustrates the effectiveness of selecting feature points that can achieve feature point matching on the standard image and images under different angles as reference points in the subsequent detection process.

The finally calculated reference feature points are used to calculate the parameters of the registration transformation model as the calculation image registration transformation model parameters of the SURF algorithm image registration, so that the image to be registered can be accurately transformed to the standard image position to achieve registration Purpose.

The SURF feature point detection algorithm is a fast and robust feature extraction algorithm based on the SIFT algorithm. The algorithm not only has good robustness to image rotation, translation, scaling and noise, but also has a good ability to deal with the invariance of illumination changes and viewing angle changes, as well as image ambiguity. At the same time, by introducing the integral image and the box filter, the operation speed is increased by about 3 times, and the comprehensive performance is more superior.

The SURF feature point detection algorithm determines the position of feature points by calculating the local extremum of the determinant of the Hessian matrix. Among them, the midpoint of image I with scale σ The Hessian matrix is defined as:

$\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; =</span>\left[\begin{array}{cc}{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; xxx</span>}}<span\; class="notranslate">\; (</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span>& {\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; xy</span>}}<span\; class="notranslate">\; (</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span>\\ {\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; xy</span>}}<span\; class="notranslate">\; (</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span>& {\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; yy</span>}}<span\; class="notranslate">\; (</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span>\end{array}\right]$

In the formula, L _xx is the second order derivative of Gaussian The result of convolution with I=(x,y), where L _xy and L _yy have the same meaning.

The SURF algorithm uses box filtering to approximate the second-order Gaussian filtering, constructs a fast Hessian matrix, and uses integral images to accelerate convolution to improve calculation speed.

Among them, as shown in FIG. 5, it is a 9×9 box filter template. In order to make the SURF feature scale-invariant, it is necessary to construct a scale space, and obtain local extremum points through the fast Hessian matrix determinant in each scale space. Here, the SURF algorithm uses box filters of different sizes to convolve with the source image to obtain feature descriptions in different scale spaces. After box filtering, the value in the x direction is denoted as D _xx , the value in the xy direction is denoted as D _xy , and the value in the y direction is denoted as D _yy . Since the box filter is an approximate estimate of the second-order Gaussian filter, the scale factor ω (ω is approximately 0.9) is introduced, and the row and column expression of the fast Hessian matrix is obtained by further solving:

ΔH＝D _xx (x)D _yy (x)-(ωD _xy (x)) ²

Among them, Δ is the response value of the box filter in the area around the image point I(x, y), and Δ is used to detect the extreme points.

After obtaining the local extreme points of the scale image in each scale space according to the fast Hessian matrix, each local extreme point and 8 adjacent points of the same scale and 9 points of its upper and lower scales constitute a 3 ×3×3 three-dimensional neighborhood. Compare each extremum point in the scale space with the 26 adjacent points in the stereo neighborhood, and only when the value of the local extremum point is larger (or smaller) than all 26 adjacent points, the local extremum points as candidate feature points. In order to perform sub-pixel positioning on candidate feature points, interpolation can be performed in scale space and image space to obtain stable feature point positions and scale values.

The quadratic fitting function used for interpolation is:

$\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; +</span>\frac{<span\; class="notranslate">\; \&PartialD;</span>{\mathrm{<span\; class="notranslate">\; D.</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; x</span>}}\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; +</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\frac{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; D.</span>}}{<span\; class="notranslate">\; \&PartialD;</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}\mathrm{<span\; class="notranslate">\; x</span>}$

The extreme point obtained by deriving the function is:

$\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; D.</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}}{<span\; class="notranslate">\; \&PartialD;</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}\frac{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; D.</span>}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; x</span>}}$

The extreme value of the function at the extreme point is:

$\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; (</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; +</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}\frac{<span\; class="notranslate">\; \&PartialD;</span>{\mathrm{<span\; class="notranslate">\; D.</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; x</span>}}\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}$

Candidate feature points with D(x)<0.03 were eliminated in the experiment.

Description of SURF feature points

In order to keep the rotation invariance of the feature points, after the position of the feature points is determined, the main direction is determined for each feature point. To this end, in a circular area with a feature point as the center and a calculated radius of 6σ (σ is the scale of the feature point), the Harr wavelet (Harr wavelet side length is 4σ) response operation is performed on the image in the x and y directions, and Use a Gaussian weighting function with a scale of 2σ to Gaussian-weight the Harr wavelet response value, so that the closer to the feature point, the greater the contribution of the response; then use the fan-shaped area of π/3 size to traverse the entire circular area, and the horizontal direction response in the domain The direction of the maximum value of the vector sum modulus and the vertical direction response is defined as the main direction of the feature point (as shown in Figure 6).

Then, with the feature point as the center, construct a square window area with a side perpendicular to the main direction and a side length of 20σ, and divide the window area into 4×4 sub-areas. In each sub-region, calculate the Harr wavelet response in the horizontal direction and vertical direction of 5σ×5σ sampling points, which are recorded as d _x and d _y respectively, and use the Gaussian weighting function with a scale of 2σ to calculate the Harr wavelet response value Gaussian weighting to increase robustness to geometric transformations. Then add the response value and the absolute value of the response value for each sub-region to form ∑d _x , ∑d _y , ∑|d _x |, ∑|d _y |. Thus, each sub-region forms a four-dimensional feature description vector V:

V＝(Σd _x ,Σd _x ,Σ|d _x |,Σ|d _y |)

For each feature point, a 4×4×4=64-dimensional feature vector is formed. In order to ensure the invariance of the illumination, the feature vector is normalized to obtain the final SURF descriptor, as shown in Figure 7.

Matching of feature points

The matching of feature points is the key to image registration, and the accuracy of matching directly affects the accuracy of subsequent registration. In the algorithm, according to the feature point neighborhood information contained in the SURF feature point descriptor, the K nearest neighbor method can be used to find out the two potential best matching points for each feature point, and the distance between the best value of the matching point and the second best The value ratio is optimized to find the best matching point pair.

Let M ₁ and M ₂ be two images I ₁ and I ₂ respectively using SURF to extract feature point sets to be matched. For any feature point m _1i in M ₁ , the Euclidean distance between M ₂ and m _1i is the smallest The two feature points are m _2j , m' _2j , and the corresponding distances are d _ij , d' _ij . If d _ij ≤α*d' _ij (experimentally α=0.65), m _1i and m _2j are considered to be the best matching point pairs.

Using the above method, all matching point pairs of the standard image and the image to be registered can be obtained Among them, S and D are the matching feature point sets of the standard image and the image to be registered respectively. Because there are certain mismatching point pairs in these matching point pairs, this paper adopts the method of triangle similarity to purify them. Randomly select three matching point pairs P ₁ and Q ₁ , P ₂ and Q ₂ , and P ₃ and Q ₃ in the feature point sets S and D to form a virtual triangle ΔP ₁ P ₂ P ₃ and ΔQ ₁ Q ₂ Q ₃ , and form a triangle pair, the lengths of the three sides are respectively set to lp ₁ , lp ₂ , lp ₃ and lq ₁ , lq ₂ , lq ₃ , if the relationship in the following formula is satisfied, it means that the selected matching point pairs are average is a valid point pair, otherwise there is a mismatch point pair, which will be eliminated to realize the purification of feature point pairs:

$\frac{{\mathrm{<span\; class="notranslate">\; lp</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}{{\mathrm{<span\; class="notranslate">\; lp</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}<span\; class="notranslate">\; \&ap;</span>\frac{{\mathrm{<span\; class="notranslate">\; lp</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; lp</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; \&ap;</span>\frac{{\mathrm{<span\; class="notranslate">\; lp</span>}}_{\mathrm{<span\; class="notranslate">\; 3</span>}}}{{\mathrm{<span\; class="notranslate">\; lp</span>}}_{\mathrm{<span\; class="notranslate">\; 3</span>}}}$

In the experiment of this embodiment, the error between the side length ratios is controlled at 0.02.

Although the embodiments and drawings of the present invention are disclosed for the purpose of illustration, those skilled in the art can understand that various replacements, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims Therefore, the scope of the present invention is not limited to what is disclosed in the embodiments and drawings.

Claims (4)

1. the high precision image matching process of a distinguished point based training, it is characterized in that: before image registration, the image subject to registration under different angles and standard picture are carried out Feature Points Matching, and extract the higher unique point of positioning precision as the reference characteristic point in actual testing process.

2. the high precision image matching process of distinguished point based training according to claim 1, it is characterized in that: the concrete selection step of described reference characteristic point is: use SURF feature point extraction algorithm extracts the unique point in two kinds of different rotary angular image respectively, and mate with the unique point in standard picture respectively, obtain the feature point set mated in standard picture, the X and Y coordinates of recording feature point; Analyze the two groups of unique point set that the match is successful obtained, choose in two unique point set and there is the on all four unique point of X and Y coordinates as reference characteristic point.

3. the high precision image matching process of distinguished point based training according to claim 1, is characterized in that: the described anglec of rotation is 2 ° to 10 °.

4. the high precision image matching process of distinguished point based training according to claim 1, it is characterized in that: described image registration adopts SURF feature point detection algorithm, using the computed image registration transformation model parameter of reference characteristic point as the image registration of SURF algorithm, image subject to registration is made to transform to standard picture position exactly.