CN103679193A - FREAK-based high-speed high-density packaging component rapid location method - Google Patents

Abstract

The invention discloses a fast positioning method for high-speed and high-density packaging components based on FREAK. The method is: using the Hessian matrix in the SURF registration method to detect the position of the key point; using the neighborhood information of the key point to determine the feature point According to the distribution of the retinal model, train and learn the comparison of sampling points, and construct the FREAK feature vector according to the comparison; according to the FREAK feature vector, use the Hamming distance as the similarity measure of the feature vector to perform nearest neighbor matching; through the matching The affine transformation equation is constructed by the feature point pairs, and the above transformation equation is solved by the least square method, and the space transformation model parameters are calculated, namely: the translation parameter m in the x direction, the translation parameter n in the y direction, and the rotation angle β. Compared with the prior art, the present invention greatly reduces the computational complexity and storage cost of feature description and matching, and realizes high-speed and high-precision sub-pixel level positioning.

Description

A fast positioning method for high-speed and high-density packaging components based on FREAK

technical field

The invention relates to the research field of identification and positioning in precision electronic assembly, in particular to a fast positioning method for high-speed and high-density packaging components based on FREAK.

Background technique

Surface mount technology originated in the United States in the 1960s, when it was mainly used in military electronics. In the 1970s, due to the vigorous development of consumer products in the Japanese electronics industry, the unparalleled manufacturing advantages of SMT gradually emerged and were rapidly promoted in the electronics industry. Since the 1980s, due to the rapid development of electronic consumer products and the full understanding of the strategic position of the electronic equipment industry in various countries, SMT, as the "fourth placement revolution", has received unprecedented attention and development. At present, SMT has affected the product level in various fields such as communications, home appliances, computers, networks, automation, aviation, aerospace, and navigation. Its related technologies and equipment are important symbols of the level of electronic information manufacturing in various countries.

Digital image registration (Image Registration) is a basic task in image processing. Its definition refers to the geometric alignment of two images of the same target captured at different times, from different perspectives or different cameras. The standard process is the process of evaluating the similarity of two or more images to determine the same name points. Image registration is a basic problem in image processing and application. It has important applications in the fields of aerial image stitching, 3D imaging, machine vision and pattern recognition, remote sensing data processing, and medical image analysis. Since the task of visual processing in SMT is consistent with that of registration, image registration technology is an important part of the SMT visual inspection system, providing the necessary preprocessing for subsequent inspections. The feature-based image registration method has become the most widely used image registration method because it does not directly depend on the gray scale, has good robustness, strong anti-interference, small amount of calculation, and fast speed. Generally speaking, the basic framework of registration includes four steps: feature detection, feature matching, transformation model parameter estimation, and image resampling.

Since Lowe and Bay proposed the SIFT and SURF algorithms respectively, the pursuit of faster and more robust feature descriptors has become a recent hot development trend. An ideal feature descriptor generally has high robustness, singularity and low algorithm complexity. In order to allow the feature description algorithm to be applied to smartphones and embedded electronic devices, Alexandre Alahi proposed a new feature point descriptor FREAK (Fast Retina Keypoint) at CVPR2012. The descriptor weighs the above three performances, and is guided by the operation speed, highlighting the real-time performance of the algorithm, and at the same time, its robustness and singularity also have good performance. The focus of FREAK registration lies in the description of features, while the work of key point positioning can use some existing common methods, such as Hessian matrix, DoH detection, Harris corner point positioning, etc. Therefore, it is of great research significance to apply the feature point descriptor FREAK to identify and locate in precision electronic assembly.

Contents of the invention

The main purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art, and provide a fast positioning method for high-speed and high-density packaging components based on FREAK. The method introduces FREAK binary features on the basis of SURF key point detection, which can greatly Improve the speed of image registration, realize high-speed and high-precision visual positioning detection of components, and have strong robustness.

The purpose of the present invention is achieved through the following technical solutions: a FREAK-based high-speed and high-density packaging component rapid positioning method, comprising the following steps:

(1) Key point positioning: Input the image to be registered and the template image, and use the Hessian matrix to detect the position of the key point;

(2) Feature description: use the neighborhood information of the key points to determine the main direction of the feature points; according to the distribution of the retina model, train and learn the comparison of sampling points, and construct the FREAK feature vector according to the comparison;

(3) Feature matching: According to the FREAK feature vector, the Hamming distance (that is, XOR operation) is used as the similarity measure of the feature vector to perform nearest neighbor matching; the affine transformation equation is constructed through the matched feature point pairs, and the least square method is used to Solve the above transformation equations to calculate the parameters of the space transformation model, namely: the translation parameter m in the x direction, the translation parameter n in the y direction, and the rotation angle β.

Specifically, the specific process of key point positioning in the step (1) is as follows:

(1-1) Input the image to be registered I(x,y) and the template image f(x,y), and generate integral images respectively;

(1-2) Construct a fast Hessian matrix, and construct a scale space based on the fast Hessian matrix; then obtain a three-dimensional scale space response map based on the integral image obtained in step (1-1);

(1-3) Perform threshold segmentation on the obtained three-dimensional scale space response map, and only keep pixels with strong response values; then, use maximum value suppression to find candidate feature points; finally, use three-dimensional quadratic fitting function Perform adjacent pixel interpolation on feature points to obtain key point positions.

Specifically, the method for generating the integral image in the step (1-1) is: for a given image Q, the point (x, y) is a certain point in the image Q, then in the integral image, the point (x, y ) is the sum of the pixel values of all pixels in the rectangular area formed by the origin of Q and (x, y) pixels. Using the integral image, the calculation of the sum of all pixel values in any rectangular area in the integral image can be simplified as accessing 4 image data and 3 addition and subtraction operations. No matter how large the area of the rectangular area is, the calculation time of the sum of pixel values in the area is constant. The advantage of this approach becomes apparent when the area is large. The SURF registration method exploits this property of integral images to guarantee almost constant computation time for image convolutions with size-variable frame filters.

Specifically, the method for constructing a fast Hessian matrix in the step (1-2) is:

For a point X(x, y) on a given image, the Hessian matrix H(x, σ) of point X at scale σ is defined as follows:

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

的卷积，L_xy(X,σ)是图像在点X处与高斯二阶偏导的卷积，L_yy(X,σ)是图像在点X处与高斯二阶偏导

的卷积；Among them, L _xx (X,σ) is the image at point X and Gaussian second-order partial derivative

The convolution of L _xy (X,σ) is the image at point X with Gaussian second-order partial derivative The convolution of L _yy (X,σ) is the image at point X with Gaussian second-order partial derivative

the convolution;

Let D _xx , D _yy , and D _xy be the results of convolution of the frame filter and the image in the x direction, y direction, and xy direction, respectively, then the approximate estimate of the Hessian determinant is:

det(H _approx )=D _xx D _yy -(0.9D _xy ) ² .

Preferably, the steps described in step (2) are as follows:

(2-1) Smooth and denoise each sampling point in the neighborhood of the key point. According to the distribution of the retinal model, different Gaussian kernel filters are used. The size of the Gaussian kernel function increases with the distance between the sampling point and the center point. index distribution;

(2-2) Determine the main direction of the feature point: select the receptive field symmetrical to the center to estimate the gradient value, and use this gradient value as the main direction of the feature point, set the point set G as the comparison set used to calculate the gradient, and the feature The point main direction O is calculated as follows:

$\mathrm{<span\; class="notranslate">\; o</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; m</span>}}\underset{{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; G</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}^{{\mathrm{<span\; class="notranslate">\; \&gamma;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}^{\mathrm{<span\; class="notranslate">\; \&gamma;</span>}\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>\frac{{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}^{{\mathrm{<span\; class="notranslate">\; \&gamma;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}^{\mathrm{<span\; class="notranslate">\; \&gamma;</span>}\mathrm{<span\; class="notranslate">\; 2</span>}}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}^{{\mathrm{<span\; class="notranslate">\; \&gamma;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}^{\mathrm{<span\; class="notranslate">\; \&gamma;</span>}\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}<span\; class="notranslate">\; ;</span>$

和

是对比感受野的中心点二维图像坐标，

和

是

和所对应的灰度值，γ₁和γ₂表示进行灰度对比的两个感受野；Among them, M is the permutation number of point set G,

and

is the two-dimensional image coordinates of the center point of the comparison receptive field,

and

yes

and The corresponding gray value, γ ₁ and γ ₂ represent two receptive fields for gray comparison;

(2-3) According to the main direction of the feature point and the distribution of the retinal model, rotate the neighborhood of the feature point, sample the retinal model of the rotated neighborhood, and generate the FREAK binary feature descriptor F according to the following formula:

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ;</span>$

α表示二值描述子的二进制左移移位值；Among them, P _α represents a pair of sampling receptive fields, and N is the descriptor length, that is, the number of receptive field pairs. If the total number of receptive fields is M, then

α represents the binary left-shift shift value of the binary descriptor;

Indicates that the image information of the previous bit in P _α is compared, and the image information used is the sum of image gray levels or the average value of the region, and γ ₁ and γ ₂ represent two receptive fields for gray level comparison;

(2-4) Obtain high-variance and non-correlated comparison pairs from existing receptive field pairs, the steps include:

(2-4-1) Create a matrix D, each row of D is a descriptor of a sampling point;

(2-4-2) Calculate the mean of each column, and the difference between the mean and 0.5 represents the variance of each column;

(2-4-3) Arrange from small to large according to the variance of each column;

(2-4-4) Keep the column with the smallest variance, and iteratively select the columns with low correlation with the reserved columns from the remaining columns until the dimension of the FREAK feature descriptor meets the predetermined requirement; then obtain the FREAK feature vector.

Preferably, the step (3) feature matching steps are as follows:

(3-1) Use Hamming distance as the similarity measure of feature vectors to perform nearest neighbor matching. The method is: first search for the description features of the first 16 bytes representing fuzzy information. If the matching distance is less than the set threshold, then get Matched feature point pairs, that is, enter step (3-2);

(3-2) For the matched pair of feature points (x ₁ , y ₁ ) and (x ₂ , y ₂ ), substitute the following affine transformation formula:

$\left[\begin{array}{c}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\\ {\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\end{array}\right]<span\; class="notranslate">\; =</span>\left[\begin{array}{c}{\mathrm{<span\; class="notranslate">\; m</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}\\ {\mathrm{<span\; class="notranslate">\; no</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}\end{array}\right]<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}\left[\begin{array}{cc}{\mathrm{<span\; class="notranslate">\; cos</span>}\mathrm{<span\; class="notranslate">\; \&beta;</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}& <span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; sin</span>}\mathrm{<span\; class="notranslate">\; \&beta;</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}\\ {\mathrm{<span\; class="notranslate">\; sin</span>}\mathrm{<span\; class="notranslate">\; \&beta;</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}& {\mathrm{<span\; class="notranslate">\; cos</span>}\mathrm{<span\; class="notranslate">\; \&beta;</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}\end{array}\right]\left[\begin{array}{c}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}\\ {\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}\end{array}\right]<span\; class="notranslate">\; ;</span>$

Solve by the least squares method to obtain the transformation parameters (m″, n″, β″, s); average the transformation parameters (m″, n″, β″) of all matched feature point pairs to obtain I(x ,y) and f(x,y) roughly transform the relationship (m,n,β).

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

1. The present invention uses the FREAK descriptor to detect the characteristics of various new surface mount components, so that the detection method has strong robustness to noise and chip offset rotation.

2. The FREAK descriptor used in the present invention has the ability to quickly describe image features. On the basis of the Hessian matrix to locate key points, it greatly reduces the computational complexity and storage cost of feature description and matching, and realizes high-speed and high-precision sub-pixel level Positioning is of great significance for the actual vision detection positioning.

Description of drawings

Fig. 1 is the flowchart of the method of the present invention.

Figure 2 is a schematic diagram of an integral image.

Figure 3 is a schematic diagram of FREAK's sampling model based on the distribution of retinal receptive fields.

FIG. 4 is a schematic diagram of sampling points selected for calculating the rotation angle.

Detailed ways

The present invention will be further described in detail below in conjunction with the embodiments and the accompanying drawings, but the embodiments of the present invention are not limited thereto.

Example 1

As shown in Figure 1, a FREAK-based fast positioning method for high-speed and high-density packaging components in this embodiment includes the following steps:

S1 key point positioning: Input the image to be registered I(x,y) and the template image f(x,y), and use the Hessian matrix to detect the clumping positions of I(x,y) and f(x,y) respectively, specifically as follows:

S1.1 Generate integral images according to images I(x,y) and f(x,y) respectively:

For a given image, such as I(x,y), point (x,y) is a point in the integral image, then the integral image value I _∑ of this point is the origin of I(x,y) and (x, y) The sum of the pixel values of all the pixel points in the rectangular area formed by the pixel points. Using the integral image, the calculation of the sum of all pixel values in any rectangular area in the integral image can be simplified as accessing 4 image data and 3 addition and subtraction operations. A schematic diagram of the integral image is shown in Figure 2. For the rectangular area shown in Figure 2, the vertices defining the rectangle are A, B, C, and D respectively, and the sum of the pixel values in this area is:

I _∑ =A+D-(C+B).

Obviously, no matter how large the area of the rectangular area is, the calculation time of the sum of pixel values in the area is constant. The advantage of this approach becomes apparent when the area is large. The SURF registration method exploits this property of integral images to guarantee almost constant computation time for image convolutions with size-variable frame filters.

S1.2 Detect feature points on I(x,y) and f(x,y) respectively, specifically including the following steps:

S1.2.1 Build a fast Hessian matrix. For a point X(x,y) on a given image, the Hessian matrix H(x,σ) of point X at scale σ is defined as follows:

$\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left[\begin{array}{cc}{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; xx</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span>& {\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; xy</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span>\\ {\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; xy</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span>& {\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; yy</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span>\end{array}\right]<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

的卷积，L_xy(X,σ)是图像在点X处与高斯二阶偏导

的卷积，L_yy(X,σ)是图像在点X处与高斯二阶偏导

的卷积；Among them, L _xx (X,σ) is the image at point X and Gaussian second-order partial derivative

The convolution of L _xy (X,σ) is the image at point X with Gaussian second-order partial derivative

The convolution of L _yy (X,σ) is the image at point X with Gaussian second-order partial derivative

the convolution;

Let D _xx , D _yy and D _xy be the results of the convolution of the frame filter and the image in the x direction, y direction and xy direction, then the approximate estimate of the Hessian determinant is:

det(H _approx )=D _xx D _yy -(0.9D _xy ) ² (2)

S1.2.2 Construct the scale space according to the fast Hessian matrix, and use the integral image obtained in step S1.1 to obtain a three-dimensional scale space response map. The construction of scale space needs to be done with the help of image pyramid. The SURF registration method is to construct a pyramid by convolving the original image with a frame filter of increasing size. Because the integral image is used to process the convolution, the calculation speed of frame filters of different sizes is the same, which improves the efficiency of the algorithm.

S1.2.3 Accurate positioning of feature points: first, threshold segmentation is performed on the obtained three-dimensional scale space response map, and only pixels with strong response values are retained; then, maximum value suppression is used to find candidate feature points; finally, use The three-dimensional quadratic fitting function performs interpolation of adjacent pixels on the feature points, so that it has the accuracy of subspace and subscale.

S2 feature description: Statize the feature information of the key point neighborhood, and generate a series of FREAK binary descriptors. Specific steps are as follows:

S2.1 Smooth and denoise each sampling point in the key point neighborhood, and use different Gaussian kernel filters according to the distribution of the retina model. As shown in Fig. 3, the size of the Gaussian kernel function is distributed exponentially with the increase of the distance between the sampling point and the center point.

S2.2 Determine the main direction of the feature points: In order to calculate the gradient of the target rotation, the FREAK algorithm selects 45 point pairs as shown in Figure 4 to calculate the angle. Let the point set G be the comparison set used to calculate the gradient. The main direction O of the feature point is calculated as follows:

$\mathrm{<span\; class="notranslate">\; o</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; m</span>}}\underset{{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; G</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}^{{\mathrm{<span\; class="notranslate">\; \&gamma;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}^{\mathrm{<span\; class="notranslate">\; \&gamma;</span>}\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>\frac{{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}^{{\mathrm{<span\; class="notranslate">\; \&gamma;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}^{\mathrm{<span\; class="notranslate">\; \&gamma;</span>}\mathrm{<span\; class="notranslate">\; 2</span>}}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}^{{\mathrm{<span\; class="notranslate">\; \&gamma;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}^{\mathrm{<span\; class="notranslate">\; \&gamma;</span>}\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}<span\; class="notranslate">\; ;</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

和

是对比感受野的中心点二维图像坐标，

和

是

和

所对应的灰度值，γ₁和γ₂表示进行灰度对比的两个感受野。Among them, M is the permutation number of point set G,

and

is the two-dimensional image coordinates of the center point of the comparison receptive field,

and

yes

and

The corresponding gray values, γ ₁ and γ ₂ represent two receptive fields for gray comparison.

S2.3 Rotate the neighborhood of the feature point according to the main direction of the feature point. Perform retinal model sampling on the rotated neighborhood, and generate the FREAK binary feature descriptor F as follows:

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ;</span>$

α表示二值描述子的二进制左移移位值；Among them, P _α represents a pair of sampling receptive fields, and its spatial distribution is shown in Figure 3; N is the descriptor length, that is, the number of receptive field pairs. If the total number of receptive fields is M, then

α represents the binary left-shift shift value of the binary descriptor;

表示对比对P_α中前面一位的图像信息，采用的图像信息为图像灰度之和或者区域均值，本实施例采用的是区域均值。γ₁和γ₂表示进行灰度对比的两个感受野。

Indicates that compared with the image information of the first bit in P _α , the image information used is the sum of image gray levels or the area average value, and this embodiment adopts the area average value. γ ₁ and γ ₂ represent two receptive fields for grayscale comparison.

S2.4 Learning to train pairs of sampling points with high variance and non-correlation. The specific steps include: (1) Create a matrix D, each row of D is a descriptor of a sampling point; (2) Calculate the mean value of each column, and the distance between the mean value and 0.5 represents the variance of each column; (3) According to each column (4) Keep the column with the smallest variance, iteratively select the column with low correlation with the reserved column from the remaining columns, until the dimension of the FREAK feature descriptor reaches the predetermined requirement.

S3 Feature Matching: Perform feature matching on feature points to obtain the spatial transformation relationship (m, n, β) between I(x,y) and f(x,y), where m and n are translations in the x and y directions respectively parameter, β is the rotation angle. Specific steps are as follows:

S3.1 Since the feature vector of the FREAK descriptor is a string of binary sequences, the Hamming distance (that is, XOR operation) can be used to replace the traditional Euclidean distance matching feature vector when measuring features.

When the FREAK algorithm is matched, it first searches for the description characteristics of the first 16 bytes representing fuzzy information. If the matching distance is less than the set threshold, the following features are matched. Such a search strategy can quickly eliminate up to 90% of irrelevant matching points.

S3.2 For the pair of matched feature points (x ₁ , y ₁ ) and (x ₂ , y ₂ ), according to the affine transformation formula

$\left[\begin{array}{c}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\\ {\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\end{array}\right]<span\; class="notranslate">\; =</span>\left[\begin{array}{c}{\mathrm{<span\; class="notranslate">\; m</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}\\ {\mathrm{<span\; class="notranslate">\; no</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}\end{array}\right]<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}\left[\begin{array}{cc}{\mathrm{<span\; class="notranslate">\; cos</span>}\mathrm{<span\; class="notranslate">\; \&beta;</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}& {<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; sin</span>}\mathrm{<span\; class="notranslate">\; \&beta;</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}\\ {\mathrm{<span\; class="notranslate">\; sin</span>}\mathrm{<span\; class="notranslate">\; \&beta;</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}& {\mathrm{<span\; class="notranslate">\; cos</span>}\mathrm{<span\; class="notranslate">\; \&beta;</span>}}^{<span\; class="notranslate">\; \&prime;</span><span\; class="notranslate">\; \&prime;</span>}\end{array}\right]\left[\begin{array}{c}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}\\ {\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}\end{array}\right]<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

The transformation parameters (m″, n″, β″, s) of formula (4) are obtained by solving by the least square method; the transformation parameters (m″, n″, β″) of all matched feature point pairs are averaged , to obtain a rough transformation relationship (m, n, β) between I(x, y) and f(x, y).

The above-mentioned embodiment is a preferred embodiment of the present invention, but the embodiment of the present invention is not limited by the above-mentioned embodiment, and any other changes, modifications, substitutions, combinations, Simplifications should be equivalent replacement methods, and all are included in the protection scope of the present invention.

Claims (7)

1. the encapsulation of the high-speed and high-density based on a FREAK components and parts method for rapidly positioning, is characterized in that, comprises the following steps:

(1) key point location: input image subject to registration and template image, adopt Hessian matrix to detect the place, position of key point;

(2) feature is described: utilize the neighborhood information of key point, determine the principal direction of unique point; According to the distribution of Modified Retinal Model, training study sampled point is to comparison, according to comparison is built to FREAK proper vector;

(3) characteristic matching: according to FREAK proper vector, adopt hamming distance as the similarity measurement of proper vector, carry out arest neighbors coupling; Unique point by coupling, to building affine transformation equation, solves above-mentioned transformation equation by least square method, calculates spatial alternation model parameter, that is: the translation parameters m of x direction is, the translation parameters n of y direction and anglec of rotation β.

2. the high-speed and high-density encapsulation components and parts method for rapidly positioning based on FREAK according to claim 1, is characterized in that, the key point location detailed process of described step (1) is as follows:

(1-1) input image I subject to registration (x, y) and template image f (x, y), and difference formation product partial image;

(1-2) build quick Hessian matrix, and build metric space according to quick Hessian matrix; Then the integral image obtaining according to step (1-1) obtains three dimension scale roomage response figure;

(1-3) in the three dimension scale roomage response figure obtaining, carry out Threshold segmentation, only retain the pixel with strong response; Then, adopt maximum value to suppress to find candidate feature point; Finally, with three-dimensional quadratic fit function, unique point is carried out to adjacent pixels interpolation, obtain key point position.

3. the high-speed and high-density based on FREAK according to claim 2 encapsulates components and parts method for rapidly positioning, it is characterized in that, in described step (1-1), the method for formation product partial image is: for given image Q, point (x, y) be in image Q certain a bit, in integral image, point (x, y) value is the pixel value sum with all pixels in the initial point of Q and the formed rectangular area of (x, y) pixel.

4. the high-speed and high-density encapsulation components and parts method for rapidly positioning based on FREAK according to claim 2, is characterized in that, the method that builds quick Hessian matrix in described step (1-2) is:

For 1 X (x, y) on Given Graph picture, put the Hessian matrix H (x, σ) of X under yardstick σ and be defined as follows:

$H(X,\mathrm{\&sigma;})=\left[\begin{array}{cc}{L}_{\mathrm{xx}}(X,\mathrm{\&sigma;})& L_{\mathrm{xy}}(X,\mathrm{\&sigma;})\\ L_{\mathrm{xy}}(X,\mathrm{\&sigma;})& L_{\mathrm{yy}}(X,\mathrm{\&sigma;})\end{array}\right];$

Wherein, L _xx(X, σ) is that image is at an X place and Gauss's second order local derviation

convolution, L _xy(X, σ) is that image is at an X place and Gauss's second order local derviation

convolution, L _yy(X, σ) is that image is at an X place and Gauss's second order local derviation

convolution;

If D _xx, D _yy, and D _xybe respectively the result of x direction, y direction and xy direction upper frame wave filter and image convolution, the approximate evaluation of Hessian determinant is:

det(H _approx)=D _xxD _yy-(0.9D _xy) ²。

5. the high-speed and high-density encapsulation components and parts method for rapidly positioning based on FREAK according to claim 1, is characterized in that, the step that described step (2) feature is described is as follows:

(2-1) each sampled point of key point neighborhood is carried out to smoothing denoising;

(2-2) determine the principal direction of unique point: select to estimate Grad with respect to centrosymmetric receptive field, and using this Grad as unique point principal direction, the collection G that sets up an office is for the contrast set for compute gradient, unique point principal direction O is calculated as follows:

$O=\frac{1}{M}\underset{P_O\&Element;G}{\mathrm{\&Sigma;}}(I\left(P_o^{{\mathrm{\&gamma;}}_1}\right)-I\left(P_o^{\mathrm{\&gamma;}2}\right))\frac{P_o^{{\mathrm{\&gamma;}}_1}-P_o^{\mathrm{\&gamma;}2}}{\left|\right|P_o^{{\mathrm{\&gamma;}}_1}-P_o^{\mathrm{\&gamma;}2}\left|\right|};$

Wherein, M is the number of permutations of point set G, with

the central point two dimensional image coordinate of contrast receptive field,

with be

with

corresponding gray-scale value, γ ₁and γ ₂represent to carry out two receptive fields of intensity contrast;

(2-3) according to the distribution of unique point principal direction and Modified Retinal Model, unique point neighborhood is rotated, postrotational neighborhood is carried out to Modified Retinal Model sampling, and press following formula generation FREAK two-value Feature Descriptor F:

$F=\underset{0<\mathrm{\&alpha;}<N}{\mathrm{\&Sigma;}}{2}^{\mathrm{\&alpha;}}T\left(P_{\mathrm{\&alpha;}}\right);$

Wherein, P _αrepresent a pair of sampling receptive field, N is descriptor length, i.e. the right number of receptive field, if receptive field total quantity is M,

α represents the binary shift left shift value of two valued description;

expression is to comparison P _αin the image information of above, the image information of employing is gradation of image sum or regional average value, γ ₁and γ ₂represent to carry out two receptive fields of intensity contrast;

(2-4) from existing receptive field centering obtain high variance and non-correlation to comparison, step comprises:

(2-4-1) create matrix D, the descriptor of a sampled point of each behavior of D;

(2-4-2) calculate the average of each row, average and 0.5 difference represent the variance of each row;

(2-4-3) according to the variance of each row, arrange from small to large;

(2-4-4) retain row of variance minimum, from remaining row, select to have with reservation row iteratively the row of low correlation, until FREAK Feature Descriptor dimension reaches pre-provisioning request; And then obtain FREAK proper vector.

6. the high-speed and high-density based on FREAK according to claim 5 encapsulates components and parts method for rapidly positioning, it is characterized in that, described step (2-1) is carried out smoothing denoising to each sampled point of key point neighborhood, according to the distribution of Modified Retinal Model, adopt different gaussian kernel filtering, the size of gaussian kernel function is along with the growth of the distance of sampled point and central point is exponential distribution.

7. the high-speed and high-density encapsulation components and parts method for rapidly positioning based on FREAK according to claim 1, is characterized in that, the step of described step (3) characteristic matching is as follows:

(3-1) adopt hamming distance as the similarity measurement of proper vector, carry out arest neighbors coupling, method is: first search represents the Expressive Features of front 16 bytes of fuzzy message, if matching distance is less than set threshold value, the unique point pair that obtains coupling, enters step (3-2);

(3-2) to coupling unique point to (x ₁, y ₁) and (x ₂, y ₂), the following affined transformation formula of substitution:

$\left[\begin{array}{c}{x}_2\\ y_2\end{array}\right]=\begin{array}{}\end{array}\left[\begin{array}{c}{m}^{\&prime;\&prime;}\\ n^{\&prime;\&prime;}\end{array}\right]+s\left[\begin{array}{cc}{\cos \mathrm{\&beta;}}^{\&prime;\&prime;}& {-\sin \mathrm{\&beta;}}^{\&prime;\&prime;}\\ {\sin \mathrm{\&beta;}}^{\&prime;\&prime;}& {\cos \mathrm{\&beta;}}^{\&prime;\&prime;}\end{array}\right]\left[\begin{array}{c}{x}_1\\ y_1\end{array}\right];$

By least square method, solve, obtain transformation parameter (m ", n ", β ", s); The transformation parameter right to the unique point of all couplings (m ", n ", β ") averages, and obtains rough transformation relation (m, n, β) between I (x, y) and f (x, y).

Images