CN104657988A - Image segmentation method for micro-fine cohesive core particles based on angular point and curvature detection - Google Patents

Abstract

An image segmentation method of cohesive ore particles based on corner detection and curvature, which is suitable for a research on images of broken mineral particles. The steps are as follows: firstly, the mineral image is preprocessed; secondly, the obtained binary image is subjected to Harris corner detection; thirdly, the concave point is identified by using the curvature information of each corner point, that is, the connection point of the cohesive particles. The characteristics of the points adopt certain criteria to determine the best segmentation path and complete the segmentation of cohesive ore particles. This method finds the corner points in the target area and combines the corner points and curvature information to identify the concave points in it. Through the directional characteristics of the concave points and the nearest neighbor criterion, the image target area is segmented, and finally the entire image is completed. The segmentation of cohesive particles in the ore particle image is simple, and can effectively segment the area of a large number of cohesive particles in the image, and restore the distribution of fine ore particles in the image to the greatest extent.

Description

Image Segmentation Method of Fine-grain Cohesive Ore Particles Based on Corner Point and Curvature Detection

technical field

The invention relates to an ore particle image segmentation method, and is especially suitable for a fine-grained ore particle image segmentation method based on corner point and curvature detection used in research on broken mineral particle images.

Background technique

The main purpose of mineral processing is to separate the useful minerals in the raw ore. One of the key steps is to grind the raw ore to achieve the purpose of dissociation of useful minerals. There is a certain relationship between the particle size of the grinding and the degree of dissociation. , so the accurate detection of ore particle size is an important technology. Generally, mineral technology uses the sieving method to detect the particle size. This method uses a limited number of sieves to measure the ore particle size, and the error is relatively large. At present, it is an accurate method to use image processing to segment and identify ore particle images. For millimeter-sized ore particles, a digital camera can be used to take pictures, but for micron-sized particles, a scanning electron microscope with higher magnification must be used. Also, due to the smaller particle size of the ore, the adhesion between particles is stronger, making the particle size The inter-adhesion phenomenon is very serious, and the particle size is an important characteristic index of mineral particle materials, and its accurate measurement has important guiding significance for the subsequent processing of particles. In order to accurately analyze the particle size characterization of mineral particles, the particle images must be segmented and separated before analysis.

Contents of the invention

Aiming at the deficiencies of the above-mentioned technologies, a method for image segmentation of fine-grained cohesive ore particles based on corner point and curvature detection with simple steps is provided.

In order to achieve the above object, the present invention adopts the following technical solutions: a method for segmenting images of fine-grained cohesive ore particles based on corner point and curvature detection, the steps of which are as follows:

a. Use a scanning electron microscope to take pictures of the ground ore particles, and use the smooth function to sequentially smooth the ore particle images, image thresholding, morphological filtering, and remove edge particles, so as to complete the binary image of the ore particles change;

b. Select the image with fine particle adhesion in the binarized ore particle image as the target area, and perform Harris corner detection on the ore particle image in the target area:

1) Use the horizontal and vertical difference operators to filter each pixel in the ore particle image of the target area, and obtain the first-order derivatives I _x and I _y of the pixels in the horizontal and vertical directions, using the formula: $m=\left[\begin{array}{cc}{I}_x^2& I_x{I}_{\mathrm{the\; y}}\\ I_x{I}_{\mathrm{the\; y}}& I_{\mathrm{the\; y}}^2\end{array}\right],$ Get the second-order Hessian matrix m of the pixel point function, where I _x I _y are the second-order partial derivatives of the pixel function, and use the Hessian matrix m to judge whether each pixel is an extreme point;

2) Utilize the discrete two-dimensional zero-mean Gaussian function formula: pair Hessian matrix $m=\left[\begin{array}{cc}{I}_x^2& I_x{I}_{\mathrm{the\; y}}\\ I_x{I}_{\mathrm{the\; y}}& I_{\mathrm{the\; y}}^2\end{array}\right]$ The four element values in are subjected to Gaussian smoothing filtering to obtain the filtered Hessian matrix m';

3) Use the formula: Substitute the filtered Hessian matrix m' value into the ore particle image pixel in the target area, and calculate the corner point cim of each pixel;

4) Compare the corner point amount cim of each pixel with the preset threshold value thresh, and mark each pixel point greater than the preset threshold value thresh. When the corner point amount cim value is greater than the preset threshold value, then judge this pixel The point is Harris corner point;

c. Take each Harris corner as the center curvature, with a radius of 5 pixels, through the formula: Calculate the circular mask corresponding to each Harris corner point, where: j represents the jth corner point, |L| is the circumference of the circular mask, |A _j | is the jth corner point as the center of the circle The mask arc of the overlapping part of the circular mask and the ore particle image of the target area;

d. Using the formula: curvature(j) > 0.5, compare and judge each corner point, the corner point conforming to the formula is the concave point P _j , and exclude the non-concave points;

e. Define the mask arc B j of the part _where the circular mask of each concave point P _j does not overlap with the ore particle image of the target area, and define the center of the mask arc B _j according to the length of the mask arc B _j Point C _j , the ray P j C _j formed from the concave point P _{j as} the emission point to the connecting midpoint C _j is the direction of the concave point P _j ;

f. Establish a list of concave point coordinates, use the coordinates of the concave point to be matched as the starting pixel point of linear segmentation, search for other concave points in the list to match, and find the concave point that satisfies the closest distance and opposite direction to the concave point to be matched as the terminal For pixel points, use the Bresenham algorithm to draw separation lines, complete the segmentation of the ore particle image in the target area, and cycle through the process of matching concave points until all the concave points in the connected domain are matched.

In step b, the method of using the Hessian matrix m to judge whether each pixel point is an extreme point is as follows: if m is a positive definite matrix, then the point is a minimum value; if m is a negative definite matrix, then this point is a maximum value , if m is an indeterminate matrix, the point is not an extremum; the corner response function R of the corner quantity cim is R=λ ₁ λ ₂ -k*(λ ₁ +λ ₂ ) ² , where: λ ₁ , λ ₂ is the two orthogonal components of the Hessian matrix m' processed by the diagonalization of the real symmetric matrix, k is the coefficient, and the value range is [0.04, 0.06]; the preset threshold thresh is filtered by the Gaussian function. The smooth contour curve generated by the four element values in the plug matrix m, the variance scale parameter of the Gaussian function, and the radius of the support domain are determined. Here, the Gaussian function scale parameter σ=2.5, the radius of the contour support domain is 1, and the value of the threshold is The interval is [0.004, 0.008]; if the number of pit points in the pit coordinate list is an odd number, the remaining pit point will be ignored after matching.

Beneficial effects: Since most of the ore particles are in a state of sharp edges and corners, the method finds the corner points existing in the target area and combines the corner points and curvature information to identify the concave points in it, and through the directional characteristics of the concave points and the The nearest neighbor criterion, so as to segment the image target area, and finally complete the segmentation of cohesive particles in the entire ore particle image. The method is simple and can effectively segment the area of a large number of cohesive particles in the image, and restore the distribution of fine ore particles in the image to the greatest extent. Condition.

Description of drawings

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

FIG. 2 is a diagram showing pit detection and pit directions of a circular mask in a target area of the present invention.

Detailed ways

Embodiments of the present invention will be further described below in conjunction with the accompanying drawings:

As shown in Figure 1, the method for image segmentation of fine grained ore particles based on corner point and curvature detection of the present invention, its steps are as follows:

a. Use a scanning electron microscope to take pictures of the ground ore particles, and use the smooth function to sequentially smooth the ore particle images, image thresholding, morphological filtering, and remove edge particles, so as to complete the binary image of the ore particles change;

b. Select the image with fine particle adhesion in the binarized ore particle image as the target area, and perform Harris corner detection on the ore particle image in the target area:

1) Use the horizontal and vertical difference operators to filter each pixel in the ore particle image of the target area, and obtain the first-order derivatives I _x and I _y of the pixels in the horizontal and vertical directions, using the formula: $m=\left[\begin{array}{cc}{I}_x^2& I_x{I}_{\mathrm{the\; y}}\\ I_x{I}_{\mathrm{the\; y}}& I_{\mathrm{the\; y}}^2\end{array}\right],$ Get the second-order Hessian matrix m of the pixel point function, where I _x Ix are the second-order partial derivatives of the pixel point function respectively, and use the Hessian matrix m to judge whether each pixel point is an extreme value point; the method of using the Hessian matrix m to judge whether each pixel point is an extreme value point is as follows: , if m is a positive definite matrix, then the point is a minimum value, if m is a negative definite matrix, then this point is a maximum value, if m is an indefinite matrix, then this point is not an extreme value;

2) Utilize the discrete two-dimensional zero-mean Gaussian function formula: pair Hessian matrix $m=\left[\begin{array}{cc}{I}_x^2& I_x{I}_{\mathrm{the\; y}}\\ I_x{I}_{\mathrm{the\; y}}& I_{\mathrm{the\; y}}^2\end{array}\right]$ The four element values in are subjected to Gaussian smoothing filtering to obtain the filtered Hessian matrix m';

3) Use the formula: Substituting the filtered Hessian matrix m' value into the ore particle image pixel points in the target area, and calculating the corner point quantity cim of each pixel point; the corner point response function R of the corner point quantity cim is R=λ ₁ λ ₂ - k*(λ ₁ +λ ₂ ) ² , where: λ ₁ and λ ₂ are the two orthogonal components of the Hessian matrix m' after diagonalization of the real symmetric matrix, k is the coefficient, and the value range is [0.04, 0.06];

4) Compare the corner point amount cim of each pixel with the preset threshold value thresh, and mark each pixel point greater than the preset threshold value thresh. When the corner point amount cim value is greater than the preset threshold value, then judge this pixel The point is the Harris corner point; the preset threshold thresh is determined by the smooth contour curve generated by the four element values in the Hessian matrix m filtered by the Gaussian function, the variance scale parameter of the Gaussian function and the radius of the support domain, where the Gaussian function scale The parameter σ=2.5, the radius of the contour support domain is 1, and the value range of the threshold is [0.004, 0.008];

c. Take each Harris corner as the center curvature, with a radius of 5 pixels, through the formula: Calculate the circular mask corresponding to each Harris corner point, where: j represents the jth corner point, |L| is the circumference of the circular mask, |A _j | is the jth corner point as the center of the circle The mask arc of the overlapping part of the circular mask and the ore particle image of the target area;

d. Using the formula: curvature(j) > 0.5, compare and judge each corner point, the corner point conforming to the formula is the concave point P _j , and exclude the non-concave points;

e. Define the mask arc B j of the part _where the circular mask of each concave point P _j does not overlap with the ore particle image of the target area, and define the center of the mask arc B _j according to the length of the mask arc B _j Point C _j , the ray P _{j C j} formed from the concave point P _j as the emission point to the midpoint C _j is the direction of the concave point _{P j ;} as shown in Figure 2, the concave point detection of the circular mask in the target area and the direction of concave points, in which P ₁ , P ₂ , and P ₃ are corner points, and A ₁ , A ₂ , and A ₃ in the figure are three circles with corner points P ₁ , P ₂ , and P ₃ as centers. The mask arc of the overlapping portion of the mask and the ore particle image of the target area, B ₁ , B ₂ , and B ₃ are the three circular masks with the corner points P ₁ , P ₂ , and P ₃ as the center of the circle and the target area. The mask arcs of the non-overlapping parts of the ore particle images, C ₁ , C ₂ , and C ₃ are the midpoints of the mask arcs B ₁ , B ₂ , and B ₃ respectively;

f. Establish a list of concave point coordinates, use the coordinates of the concave point to be matched as the starting pixel point of linear segmentation, search for other concave points in the list to match, and find the concave point that satisfies the closest distance and opposite direction to the concave point to be matched as the terminal Pixels, use the Bresenham algorithm to draw the separation line, complete the segmentation of the ore particle image in the target area, and cycle the process of matching concave points until all the concave points in the connected domain are matched. If the number of concave points in the concave point coordinate list is Odd number, after matching, ignore the remaining concave point.

Claims (5)

1., based on a microfine adhesion ore particles image partition method for angle point and curvature measuring, it is characterized in that step is as follows:

A. scanning electron microscope is used to get figure to ground ore particles, utilize smooth function to the step of the smoothing process of ore particles image sequence, image threshold, shape filtering, removal edge particle, thus complete the binaryzation to ore particles image;

B. the image that there is microfine adhesion situation in the ore particles image after binaryzation is selected and is used as target area, Harris Corner Detection is carried out to the ore particles image of target area:

1) utilize level, vertically difference operator to carry out filtering to each pixel in the ore particles image of target area, obtain obtaining the first order derivative I of pixel in horizontal and vertical direction _x, I _y, utilize formula: $m=\left[\begin{array}{cc}{I}_x^2& I_x{I}_y\\ I_x{I}_y& I_y^2\end{array}\right],$ Obtain the second order Hessian matrix m of pixel function, in formula i _x ², I _y ², I _xi _ybe respectively the second-order partial differential coefficient of this pixel function, and whether each pixel is extreme point to utilize Hessian matrix m to judge;

2) discrete two-dimensional zero-mean gaussian function formula is utilized: to Hessian matrix $m=\left[\begin{array}{cc}{I}_x^2& I_x{I}_y\\ I_x{I}_y& I_y^2\end{array}\right]$ In four element values carry out Gaussian smoothing filter, obtain filtered Hessian matrix m';

3) formula is utilized: filtered Hessian matrix m' value is substituted into the ore particles image slices vegetarian refreshments of target area, calculate the angle point amount cim of each pixel;

4) the angle point amount cim of each pixel and pre-set threshold value thresh is compared, mark the pixel that each is greater than pre-set threshold value thresh, when angle point amount cim value is greater than pre-set threshold value, then judge that this pixel is Harris angle point;

C. using each Harris angle point as center of circle curvature, radius is 5 pixels, passes through formula: calculate the circular mask that each Harris angle point is corresponding, in formula: j represents a jth angle point, | L| is the girth of circular mask, | A _j| for a jth angle point be the circular mask in the center of circle and the ore particles image of target area coincide part mask camber line;

D. utilize formula: curvature (j) > 0.5, compare judgement to each angle point, the angle point of coincidence formula is concave point P _j, and get rid of non-concave point;

E. each concave point P is defined _jcircular mask and the mask camber line B of ore particles image not intersection of target area _j, according to mask camber line B _jlength definition mask camber line B _jmid point C _j, with concave point P _jfor launching site is to connection mid point C _jthe ray P formed _jc _jbe this concave point P _jdirection;

F. concave point list of coordinates is set up, using the top pixel of concave point coordinate to be matched as linear partition, other concave points in search listing mate with it, find and meet and concave point that direction contrary nearest with concave point to be matched as terminal pixel, Bresenham algorithm is adopted to draw defiber, complete the segmentation of the ore particles image of target area, the process of circulation coupling concave point, until concave points all in this connected domain is all mated.

2. the microfine adhesion ore particles image partition method based on angle point and curvature measuring according to claim 1, its feature exists: whether each pixel is the method for extreme point and is to utilize Hessian matrix m to judge, if m is positive definite matrix, then this point is minimal value, if m is negative definite matrix, then this point is maximum value, if m is indefinite matrix, then this point is not extreme value.

3. the microfine adhesion ore particles image partition method based on angle point and curvature measuring according to claim 1, its feature exists: the angle point response function R of angle point amount cim is R=λ ₁λ ₂-k* (λ ₁+ λ ₂) ², in formula: λ ₁, λ ₂for two quadrature components that Hessian matrix m' obtains after real symmetric matrix diagonalization process, k is coefficient, and span is [0.04,0.06].

4. the microfine adhesion ore particles image partition method based on angle point and curvature measuring according to claim 1, its feature exists: the radius of the level and smooth contour curve that pre-set threshold value thresh is produced by four element values in Gaussian function filtered Hessian matrix m, the variance measure parameter of Gaussian function and supporting domain determines, here Gaussian function scale parameter σ=2.5, the radius in skeletal support territory is 1, then the interval of threshold value is [0.004,0.008].

5. the microfine adhesion ore particles image partition method based on angle point and curvature measuring according to claim 1, its feature exists: if the quantity of concave point is odd number in concave point list of coordinates, then after overmatching, ignores a remaining concave point.