Ore grade identification method and system based on color three-dimensional reconstruction

Abstract

The invention belongs to the technical field of ore grade identification, discloses an ore grade identification method and system based on color three-dimensional reconstruction, designs a mineral three-dimensional reconstruction system based on multi-view vision, and establishes a surface dynamic stereoscopic vision three-dimensional reconstruction method based on line structured light. Comprising joint calibration of a camera and a laser ray generator, relative attitude estimation between adjacent frames and global optimization of an absolute attitude, and the limitation of a static stereoscopic vision measurement range is overcome. And carrying out three-dimensional reconstruction test and analysis on various mineral models, and then carrying out grade identification on the mineral models by utilizing the reconstructed models. The three-dimensional reconstruction system realizes non-contact, high precision and full coverage of minerals. According to the mineral three-dimensional reconstruction system based on the multi-view vision, mineral image features are increased through a three-dimensional reconstruction method, the precision of grade recognition of minerals is improved, the system carries out a non-contact full-coverage mode on the minerals in the whole process, and safety and high efficiency of the process are ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of ore grade identification, and in particular relates to an ore grade identification method and system based on color three-dimensional reconstruction. Background Art

[0002] Ore grade identification is mainly based on laboratory testing methods. Although this method has irreplaceable effects in terms of precision and accuracy, it is time-sensitive and difficult to operate, making it difficult to meet the needs of rapid grade acquisition at iron ore construction sites. Ore grade identification is the extraction and identification of shallow information from images under the same background. In order to improve the accuracy of mineral texture restoration and enhance grade identification efficiency, three-dimensional reconstruction can combine multiple images into one, increasing image features.

[0003] Through the above analysis, the problems and defects of the prior art are as follows:

[0004] Ore grade identification is mainly based on laboratory testing methods. Although this method has irreplaceable effects in terms of precision and accuracy, it has poor timeliness and is difficult to operate, making it difficult to meet the needs of rapid grade acquisition at iron ore construction sites. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention provides a method and system for identifying ore grade based on color three-dimensional reconstruction.

[0006] The present invention is implemented in this way: a method for identifying ore grade based on color three-dimensional reconstruction comprises:

[0007] Step 1, turn on the motor and adjust the speed regulator to make the vibrating screen and belt conveyor move;

[0008] Step 2: The ore falls to the tail end of the belt conveyor, the camera takes a picture, and the image is sent to the 3D reconstruction system;

[0009] Step 3, perform three-dimensional reconstruction on the input image;

[0010] Step 4: The 3D reconstructed image is passed to the intelligent recognition system for grade recognition based on the MobileNet-SSA algorithm.

[0011] Further, the three-dimensional reconstruction of the input image is performed; the specific steps are as follows:

[0012] First, each binocular camera is composed of a laser line generator, which moves along the horizontal push-sweep direction and calculates the three-dimensional coordinates of the center of the line structured light stripe at each position; as long as the three-dimensional coordinates of the stripe center at each position are unified into the reference coordinate system, the three-dimensional point cloud of the entire measured surface can be obtained; the intrinsic parameter matrix of one camera is defined as R and the other cameras as R′, and the rotation and translation of the other cameras (expressed as Cr ) relative to a fixed camera (denoted by C l ) as L and t respectively; Taking the left and right cameras as an example, let the line structured light stripe at a position of the three-dimensional coordinate P of the definition point in the left camera coordinate system be P(X W , Y W , Z W ), and the projections P of the point on the left and right image planes are p(u, v), p′(u′, v′). It must be emphasized that the projection points mentioned are undistorted image coordinates; according to the perspective projection imaging model of the camera, we have:

[0013]

[0014] β, μ are proportional factors, (f x , f y ) and (f′ x , f′ y ) is the focal length of the left and right cameras in the x and y directions, and (c x , c y ), (c′ x , c′ y ) are the principal points of the left and right cameras respectively;

[0015] The normal vector of the line structured light plane is defined as n = [ABCD] T , then we have:

[0016] AX W +BY W +CZ W +D=0 (3)

[0017] The perspective projection imaging model of the left camera and the plane equation of the line structured light are used to calculate the 3D coordinates of the center of the line structured light stripes, avoiding the matching of the line structured light stripes on the image planes of the left and right cameras during the 3D reconstruction process; the 3D coordinates of the center of the stripes can be calculated by the integral equation; the formula is as follows:

[0018]

[0019] Secondly, the joint calibration of the single camera and the laser line generator is as follows:

[0020] Let the physical 3D coordinates of a chessboard corner in the corresponding chessboard coordinate system be defined as ( X W , Y W , Z W ), where Z W = 0; the projection of the chessboard angle on the camera image plane is expressed as g(u g ,v g), according to the perspective projection imaging model, we get the equation:

[0021]

[0022] Define H = K[L c t c ], and h=[h 11 ,h 12 ,h 13 ,h 14 ,h 21 ,h 22 ,h 23 ,h 24 ,h 31 ,h 32 ,h 33 ,h 34 ] T ,h ij represents the element in row i and column j in H;

[0023] Eliminate the ratio factor β and perform singular value decomposition:

[0024] sR[r1 r2 t c ]=[h1 h2 h4] (6)

[0025] r i and h i Respectively represent L c The i-th column of and the i-th column of H;

[0026] Then L c and t c The rough value of can be calculated by the following formula:

[0027]

[0028] In order to obtain accurate r-process and matching, the minimization function of pixel reprojection error is established, as shown in equation (8); the LM iterative algorithm is used to calculate the refined solution;

[0029]

[0030] where x′ is the mapping function of the image coordinates of the chessboard corners, and x j is the real image coordinate of the jth chessboard corner, X j is the physical three-dimensional coordinate of the jth chessboard corner in the chessboard coordinate system;

[0031] After the rotation and translation of the chessboard plane relative to the camera coordinate system are determined, the three-dimensional coordinates of any point on the projected structured light in the chessboard coordinate system can be calculated by the following formula:

[0032]

[0033] (u s , v s ) is the image coordinate of a point on the projected structured light; finally, the three-dimensional coordinates of any point on the projected structured light in the camera coordinate system can be determined by the following formula;

[0034]

[0035] The homogeneous coordinates of any two points on the projection line lAB in the camera coordinate system are defined as M a =[x c1 y c1 z c1 1] T , M b =[x c2 y c2 z c2 1] T ; Equation 11 is used to represent the projection line;

[0036] L=M a M b T -M b M a T (11)

[0037] Place the chessboard in different orientations and calculate the corresponding Plucker matrix of the projected structured light, resulting in the overdetermined linear equation can be stacked vertically by all Plucker matrices; similarly, the normal vector n of the line structured light plane is finally calculated by singular value decomposition;

[0038] Then the relative pose estimation between adjacent frames is performed; the process is as follows:

[0039]

[0040] The rotation and translation of the left camera at the jth position relative to the left camera at the ith position are defined as and A pair of points with the same name obtained by the left and right cameras are defined as {P iL (u iL ,v iL ),P iR (u iR ,v iR )} and the j-th position {P jL (u jL ,v jL ),P jR (u jR ,v jR)}; Since the internal and external parameters of the binocular module are known, the three-dimensional coordinates of the same-name points at the i-th and j-th positions can be calculated by the triangulation method of binocular stereo vision, which are expressed as (x wi ,y wi ,z wi ) and (x wj ,y wj ,z wj ); When calculating the 3D coordinates of each position of the binocular module, the left camera coordinate system of the corresponding position is regarded as the world coordinate system; according to the perspective projection imaging model of the camera, it can be obtained:

[0041] A minimization function of the reprojection error is constructed in the object space to calculate the accurate relative pose between adjacent frames, as shown in Equation (14);

[0042]

[0043] are the three-dimensional coordinates of the kth feature point at the ith and jth positions respectively;

[0044] Finally, the absolute posture is globally optimized, and the absolute posture of the binocular module at each position is calculated using a global optimization algorithm.

[0045] The absolute posture of the binocular module at the i-th and j-th positions is defined as L i , T i and L j , T j , P i and P j Represent the left camera coordinate system of the i-th and j-th positions respectively;

[0046]

[0047] again have:

[0048]

[0049] in Respectively represent L i and L j The σ-th column vector of ;

[0050] When there is a relative pose between the binocular module at the initial position and the binocular module at the i-th or j-th position, there is:

[0051]

[0052] make A is the identity matrix;

[0053]

[0054] After obtaining all relative poses between adjacent frames, A can be obtained based on equation (18): r x r =b r ; A r is formed by vertically stacking the coefficient matrices of equation (18), x r is the combination of the column vectors of the stereo module’s rotation matrix at all positions relative to the reference coordinate system, b r The vectors on the right side of the equation are stacked vertically; according to the least squares method x r =(A r T A r ) -1 A r T b r ;

[0055]

[0056]

[0057] Re-estimate the rotation matrix by using the forced orthogonality method;

[0058] L i is the rotation matrix calculated by the least squares algorithm described above, is the corresponding forced orthogonal matrix;

[0059] According to equation (15), we can obtain:

[0060]

[0061] Similar to equation (18), the global optimal solution for the translation of the binocular module relative to the reference coordinate system can also be obtained based on the least squares principle.

[0062] Furthermore, the depth-separable convolution structure of the MobileNet-SSA algorithm is divided into three parts:

[0063] (1) Channel-by-channel convolution: conventional convolution has three convolution kernels responsible for one common channel, while channel-by-channel convolution (DW convolution) has one convolution kernel responsible for one channel. DW convolution is performed in a two-dimensional plane. The number of convolution kernels is the same as the number of channels and feature maps in the previous layer, which ensures the depth of the network and the extraction of information while reducing parameters.

[0064] (2) Point-by-point convolution: The operation of point-by-point convolution (PW convolution) is very similar to the conventional convolution operation. The size of the convolution kernel is 1x1xM, where M is the number of channels in the previous layer. Therefore, the convolution operation here will perform a weighted combination of the maps in the previous step in the depth direction to generate a new feature map. There are as many feature map outputs as there are convolution kernels.

[0065] (3) Calculation of computational effort: The computational effort of the model is referred to as computational effort, which is a key indicator for measuring the algorithm's running speed and training speed. In the example, the input image is 3 (M) channels and 4 channels, and the final output is 4 (N) feature maps. The length and width of the image are DF, and the convolution kernel size is (3x3)DKxDK.

[0066] Another object of the present invention is to provide an ore grade identification system based on color three-dimensional reconstruction, comprising:

[0067] The regulating module is used to turn on the motor and adjust the speed regulator to make the vibrating screen and belt conveyor move;

[0068] The camera module is used for the ore to fall to the tail end of the belt conveyor. The camera takes pictures and sends the images to the 3D reconstruction system.

[0069] A three-dimensional reconstruction module, used for performing three-dimensional reconstruction of the input image;

[0070] The grade recognition module is used to transmit the 3D reconstructed image to the intelligent recognition system for grade recognition based on the MobileNet-SSA algorithm.

[0071] Another object of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the ore grade identification method based on color three-dimensional reconstruction.

[0072] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the ore grade identification method based on color three-dimensional reconstruction.

[0073] Another object of the present invention is to provide an information data processing terminal, which is used to implement the ore grade identification system based on color three-dimensional reconstruction.

[0074] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0075] First, the present invention designs a mineral 3D reconstruction system based on multi-eye vision, and establishes a surface dynamic stereoscopic vision 3D reconstruction method based on line structured light, including joint calibration of camera and laser line generator, relative posture estimation between adjacent frames and global optimization of absolute posture, which overcomes the limitation of static stereoscopic vision measurement range. 3D reconstruction tests and analysis are carried out on various mineral models, and then the grade is identified by using their reconstructed models. The 3D reconstruction system realizes non-contact, high-precision and full coverage of minerals.

[0076] The multi-vision-based mineral 3D reconstruction system of the present invention increases the mineral image features by the 3D reconstruction method, improves the accuracy of mineral grade identification, and implements a contactless full coverage mode for the minerals throughout the process, ensuring the safety and efficiency of the process. Moreover, this 3D reconstruction method overcomes the limitations of the static stereoscopic vision measurement range and plays a key role in improving the overall productivity and efficiency of mining operations, thereby promoting industrial growth and operational efficiency.

[0077] Second, the technical problems solved by the present invention in industrial applications and the significant technical progress

[0078] 1. Existing technical problems solved

[0079] 1. Insufficient accuracy in ore grade identification

[0080] Most existing technologies rely on traditional ore grade identification methods, such as spectral analysis or manual detection. Due to limitations in instrument accuracy or human errors, it is difficult to achieve efficient and accurate grade determination.

[0081] The present invention combines color 3D reconstruction technology with the MobileNet-SSA algorithm, starting from the multi-dimensional features of the ore such as shape, color and surface texture, and significantly improves the accuracy of grade identification (reaching 98%).

[0082] 2. Data collection and processing in dynamic scenes is difficult

[0083] Ore grade detection mostly occurs during dynamic transportation. Traditional methods are unable to capture the complete characteristics of the ore under high-speed movement conditions, resulting in incomplete or distorted data collection.

[0084] The present invention adopts a binocular camera and a dynamic three-dimensional reconstruction method to achieve accurate shooting and three-dimensional modeling of high-speed falling ore, ensuring the integrity and reliability of the data.

[0085] 3. Lack of intelligent learning and optimization mechanism

[0086] Existing recognition systems are usually static models that cannot be optimized based on new data and changes in the usage environment, resulting in poor adaptability of the recognition model.

[0087] The present invention introduces an autonomous learning module, which dynamically optimizes the taste recognition model through user feedback and real-time data training, so that the system can maintain the optimal state in different scenarios.

[0088] 4. Low efficiency in industrial applications

[0089] The traditional ore testing process is cumbersome, involving a variety of independent equipment and a low degree of automation, making it difficult to meet the modern mining industry's demand for efficient testing.

[0090] The present invention uses a fully automated ore grade identification system that integrates a vibrating screen, a belt conveyor, a binocular camera, a three-dimensional reconstruction system and an intelligent identification system, thereby greatly improving detection efficiency.

[0091] 2. Significant Technological Advances

[0092] 1. Comprehensively improve the accuracy and efficiency of grade recognition

[0093] The present invention combines dynamic 3D reconstruction technology with the lightweight MobileNet-SSA algorithm, which significantly reduces computational complexity and hardware cost while ensuring high recognition accuracy.

[0094] The system's detection speed is 6 times that of traditional methods, suitable for high-speed conveyor line scenarios and meeting the needs of large-scale industrial applications.

[0095] 2. Realize automation and intelligence of ore detection

[0096] The system automates the entire process of ore transportation, photography, 3D modeling, grade determination and result feedback, realizing intelligent operations from data collection to result output, reducing manual participation and error rates.

[0097] The introduction of the autonomous learning module enables the system to continuously optimize itself, enhancing its adaptability to new ore varieties and complex scenarios.

[0098] 3. High-precision modeling for dynamic environments

[0099] The dynamic stereoscopic vision three-dimensional reconstruction method of the present invention solves the difficult problem of high-speed moving ore modeling by optimizing the algorithm, and ensures the modeling accuracy and real-time performance.

[0100] The system has a fixed image resolution of 256×256, which reduces computing power requirements while ensuring accurate extraction of key features.

[0101] 4. Promote the intelligent upgrading of mining detection technology

[0102] The technology of the present invention is not only applicable to ore grade detection, but also provides a reference for intelligent detection technology for other industrial fields (such as material surface detection, product quality control, etc.).

[0103] The integration and versatility of the system have promoted the transformation of mining detection from traditional methods to intelligent technology and improved the level of automation in the industry.

[0104] 3. Industrialization Value

[0105] 1. Improve production efficiency

[0106] The system realizes the real-time linkage between ore transportation and grade detection, which improves the detection efficiency by more than 3 times and can significantly reduce the operating costs of mining companies.

[0107] 2. Optimize resource allocation

[0108] Accurate grade identification results help to classify and efficiently utilize ore, and provide technical support for optimal resource allocation and maximization of economic benefits.

[0109] 3. Adapt to multiple scenarios

[0110] The modular design of the system (vibrating screen, belt conveyor, 3D reconstruction module and identification system) facilitates deployment in different mining production lines and adapts to diverse needs such as ore types and transportation methods.

[0111] The present invention provides an intelligent and automated solution to the pain points of low accuracy, low efficiency, and poor adaptability in ore grade identification. Through dynamic 3D reconstruction technology and improved MobileNet-SSA algorithm, the present invention has achieved significant technological breakthroughs in data acquisition, 3D modeling, and grade identification, providing important support for the intelligent development of mining detection and has broad industrial application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0112] Figure 1 It is a flow chart of an ore grade identification method based on color three-dimensional reconstruction provided by an embodiment of the present invention.

[0113] Figure 2 It is a structural block diagram of an ore grade identification system based on color three-dimensional reconstruction provided by an embodiment of the present invention.

[0114] Figure 3 It is a schematic diagram of a color three-dimensionally reconstructed ore grade identification system provided by an embodiment of the present invention.

[0115] Figure 4 It is a flow chart of a color three-dimensionally reconstructed ore grade identification system provided by an embodiment of the present invention.

[0116] Figure 5 This is a workflow diagram of a color three-dimensionally reconstructed ore grade identification system provided by an embodiment of the present invention.

[0117] Figure 6It is a principle flow chart of the ore grade identification system provided by an embodiment of the present invention.

[0118] Figure 7 is a structural diagram of a camera provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0119] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0120] like Figure 1 As shown, an ore grade identification method based on color 3D reconstruction provided by an embodiment of the present invention comprises the following steps:

[0121] S101, turn on the motor and adjust the speed regulator to make the vibrating screen and belt conveyor move;

[0122] S102, the ore falls to the tail end of the belt conveyor, the camera takes a picture, and the image is sent to the 3D reconstruction system;

[0123] S103, performing three-dimensional reconstruction on the input image;

[0124] S104, the 3D reconstructed image is transmitted to the intelligent recognition system for grade recognition based on the MobileNet-SSA algorithm.

[0125] The embodiment of the present invention provides a method for performing three-dimensional reconstruction of an input image; the specific steps are as follows:

[0126] First, each binocular camera is composed of a laser line generator, which moves along the horizontal push-sweep direction and calculates the three-dimensional coordinates of the center of the line structured light stripe at each position; as long as the three-dimensional coordinates of the stripe center at each position are unified into the reference coordinate system, the three-dimensional point cloud of the entire measured surface can be obtained; the intrinsic parameter matrix of one camera is defined as R and the other cameras as R′, and the rotation and translation of the other cameras (expressed as C r ) relative to a fixed camera (denoted by C l ) as L and t respectively; Taking the left and right cameras as an example, let the line structured light stripe at a position of the three-dimensional coordinate P of the definition point in the left camera coordinate system be P(X W , Y W , Z W ), and the projections P of the point on the left and right image planes are p(u, v), p′(u′, v′). It must be emphasized that the projection points mentioned are undistorted image coordinates; according to the perspective projection imaging model of the camera, we have:

[0127]

[0128] β, μ are proportional factors, (f x , f y ) and (f′ x , f′ y ) is the focal length of the left and right cameras in the x and y directions, and (c x , c y ), (c′ x , c′ y ) are the principal points of the left and right cameras respectively;

[0129] The normal vector of the line structured light plane is defined as n = [ABCD] T , then we have:

[0130] AX W +BY W +CZ W +D=0 (3)

[0131] The perspective projection imaging model of the left camera and the plane equation of the line structured light are used to calculate the 3D coordinates of the center of the line structured light stripes, avoiding the matching of the line structured light stripes on the image planes of the left and right cameras during the 3D reconstruction process; the 3D coordinates of the center of the stripes can be calculated by the integral equation; the formula is as follows:

[0132]

[0133] Secondly, the joint calibration of the single camera and the laser line generator is as follows:

[0134] Let the physical three-dimensional coordinates of a chessboard corner in the corresponding chessboard coordinate system be defined as (X W , Y W , Z W ), where Z W = 0; the projection of the chessboard angle on the camera image plane is expressed as g(u g ,v g ), according to the perspective projection imaging model, we get the equation:

[0135]

[0136] Define H = K[L c t c ], and h=[h 11 ,h 12 ,h 13 ,h 14 ,h 21 ,h 22 ,h 23 ,h 24 ,h 31 ,h 32 ,h 33 ,h34 ] T ,h ij represents the element in row i and column j in H;

[0137] Eliminate the ratio factor β and perform singular value decomposition:

[0138] sR[r1 r2 t c ]=[h1 h2 h4] (6)

[0139] r i and h i Respectively represent L c The i-th column of and the i-th column of H;

[0140] Then L c and t c The rough value of can be calculated by the following formula:

[0141]

[0142] In order to obtain accurate r-process and matching, the minimization function of pixel reprojection error is established, as shown in equation (8); the LM iterative algorithm is used to calculate the refined solution;

[0143]

[0144] where x′ is the mapping function of the image coordinates of the chessboard corners, and x j is the real image coordinate of the jth chessboard corner, X j is the physical three-dimensional coordinate of the jth chessboard corner in the chessboard coordinate system;

[0145] After the rotation and translation of the chessboard plane relative to the camera coordinate system are determined, the three-dimensional coordinates of any point on the projected structured light in the chessboard coordinate system can be calculated by the following formula:

[0146]

[0147] (u s , v s ) is the image coordinate of a point on the projected structured light; finally, the three-dimensional coordinates of any point on the projected structured light in the camera coordinate system can be determined by the following formula;

[0148]

[0149] The homogeneous coordinates of any two points on the projection line lAB in the camera coordinate system are defined as M a =[x c1 y c1 z c1 1] T , M b =[xc2 y c2 z c2 1] T ; Equation 11 is used to represent the projection line;

[0150] L=M a M b T -M b M a T (11)

[0151] Place the chessboard in different orientations and calculate the corresponding Plucker matrix of the projected structured light, resulting in the overdetermined linear equation can be stacked vertically by all Plucker matrices; similarly, the normal vector n of the line structured light plane is finally calculated by singular value decomposition;

[0152] Then the relative pose estimation between adjacent frames is performed; the process is as follows:

[0153]

[0154] The rotation and translation of the left camera at the jth position relative to the left camera at the ith position are defined as and A pair of points with the same name obtained by the left and right cameras are defined as {P iL (u iL ,v iL ),P iR (u iR ,v iR )} and the j-th position {P jL (u jL ,v jL ),P jR (u jR ,v jR )}; Since the internal and external parameters of the binocular module are known, the three-dimensional coordinates of the same-name points at the i-th and j-th positions can be calculated by the triangulation method of binocular stereo vision, which are expressed as (x wi ,y wi ,z wi ) and (x wj ,y wj ,z wj ); When calculating the 3D coordinates of each position of the binocular module, the left camera coordinate system of the corresponding position is regarded as the world coordinate system; according to the perspective projection imaging model of the camera, it can be obtained:

[0155] A minimization function of the reprojection error is constructed in the object space to calculate the accurate relative pose between adjacent frames, as shown in Equation (14);

[0156]

[0157] are the three-dimensional coordinates of the kth feature point at the ith and jth positions respectively;

[0158] Finally, the absolute posture is globally optimized, and the absolute posture of the binocular module at each position is calculated using a global optimization algorithm.

[0159] The absolute posture of the binocular module at the i-th and j-th positions is defined as L i , T i and L j , T j , P i and P j Represent the left camera coordinate system of the i-th and j-th positions respectively;

[0160]

[0161] again have:

[0162]

[0163] in Respectively represent L i and L j The σ-th column vector of ;

[0164] When there is a relative pose between the binocular module at the initial position and the binocular module at the i-th or j-th position, there is:

[0165]

[0166] make A is the identity matrix;

[0167]

[0168] After obtaining all relative poses between adjacent frames, A can be obtained based on equation (18): r x r =b r ; A r is formed by vertically stacking the coefficient matrices of equation (18), x r is the combination of the column vectors of the stereo module’s rotation matrix at all positions relative to the reference coordinate system, b r The vectors on the right side of the equation are stacked vertically; according to the least squares method x r =(A r T A r ) -1 Ar T b r ;

[0169]

[0170] Re-estimate the rotation matrix by using the forced orthogonality method;

[0171] L i is the rotation matrix calculated by the least squares algorithm described above, is the corresponding forced orthogonal matrix;

[0172] According to equation (15), we can obtain:

[0173]

[0174] Similar to equation (18), the global optimal solution for the translation of the binocular module relative to the reference coordinate system can also be obtained based on the least squares principle.

[0175] The depth-separable convolution structure of the MobileNet-SSA algorithm provided by the embodiment of the present invention is divided into three parts:

[0176] (1) Channel-by-channel convolution: conventional convolution has three convolution kernels responsible for one common channel, while channel-by-channel convolution (DW convolution) has one convolution kernel responsible for one channel. DW convolution is performed in a two-dimensional plane. The number of convolution kernels is the same as the number of channels and feature maps in the previous layer, which ensures the depth of the network and the extraction of information while reducing parameters.

[0177] (2) Point-by-point convolution: The operation of point-by-point convolution (PW convolution) is very similar to the conventional convolution operation. The size of the convolution kernel is 1x1xM, where M is the number of channels in the previous layer. Therefore, the convolution operation here will perform a weighted combination of the maps in the previous step in the depth direction to generate a new feature map. There are as many feature map outputs as there are convolution kernels.

[0178] (3) Calculation of computational effort: The computational effort of the model is referred to as computational effort, which is a key indicator for measuring the algorithm's running speed and training speed. In the example, the input image is 3 (M) channels and 4 channels, and the final output is 4 (N) feature maps. The length and width of the image are DF, and the convolution kernel size is (3x3)DKxDK.

[0179] like Figure 2 As shown, an ore grade identification system based on color 3D reconstruction provided by an embodiment of the present invention includes:

[0180] The regulating module is used to turn on the motor and adjust the speed regulator to make the vibrating screen and belt conveyor move;

[0181] The camera module is used for the ore to fall to the tail end of the belt conveyor. The camera takes pictures and sends the images to the 3D reconstruction system.

[0182] A three-dimensional reconstruction module, used for performing three-dimensional reconstruction of the input image;

[0183] The grade recognition module is used to transmit the 3D reconstructed image to the intelligent recognition system for grade recognition based on the MobileNet-SSA algorithm.

[0184] Another object of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the ore grade identification method based on color three-dimensional reconstruction.

[0185] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the ore grade identification method based on color three-dimensional reconstruction.

[0186] Another object of the present invention is to provide an information data processing terminal, which is used to implement the ore grade identification system based on color three-dimensional reconstruction.

[0187] The present invention is specifically implemented:

[0188] The purpose of the present invention is to provide an ore grade identification system based on color three-dimensional reconstruction.

[0189] To achieve the above objectives, the technical solution of the present invention is as follows: an ore grade identification system based on color three-dimensional reconstruction consists of a motor, a vibrating screen, a belt conveyor, a speed regulator, a binocular camera and a frame, a three-dimensional reconstruction system, and an intelligent identification system.

[0190] Motor: Provides power for belt conveyor and vibrating screen.

[0191] Vibrating screen: used to store ore and send it to the belt conveyor.

[0192] Conveyor belt: used to transport ore to the camera location.

[0193] Binocular camera: used to take pictures of falling ore.

[0194] The 3D reconstruction system includes: a global optimized dynamic stereoscopic vision 3D reconstruction method based on online structured light (the image resolution is fixed at 256x256).

[0195] Intelligent recognition system includes:

[0196] Knowledge acquisition module: acquire relevant professional knowledge.

[0197] Sub-database: stores image information of different minerals or the same minerals.

[0198] Autonomous learning module: stores the obtained three-dimensional reconstructed image and stores the data for the next optimization of the mineral grade recognition model.

[0199] Knowledge base: used to store professional knowledge.

[0200] Inference engine: Based on the relevant knowledge stored in the knowledge base, it integrates the detection data for calculation and then performs grade identification.

[0201] The method for realizing ore grade identification by color 3D reconstruction using the above system includes the following steps (the work flow diagram of the ore grade identification system by color 3D reconstruction is as follows: Figure 5 (shown)

[0202] The first recognition module of the Mobile Net-SSA algorithm is the SSA algorithm built using the improved MobileNet. The MobileNet model is a lightweight neural network model developed by Google, which is mainly used in various embedded devices. Its core is the deep separable convolution block. The SSA algorithm combines the advantages of the two mainstream algorithms Fast-RCNN and YOLO. Its detection speed is 6 times that of Fast-RCNN, and its map value is 11% higher than that of YOLO.

[0203] The Mobile Net-SSA algorithm uses an improved MobileNet network to replace the VGG16 network in SSA. This can improve the model's running speed and reduce the size of the model while keeping other structures unchanged. Combined with the located ore position, the mask information is used to extract the ore center image, and the KNN model is used for secondary prediction. The two prediction results are combined for weighted processing to output the final result.

[0204] MobileNet is a lightweight convolutional neural network focused on embedded devices. Compared with traditional convolutional neural networks, the use of deep separable convolution can greatly reduce the model parameters and computational complexity while slightly reducing the model accuracy. The deep separable convolution structure in MobileNet is divided into three parts. (1) Channel-by-channel convolution. Conventional convolution has three convolution kernels responsible for one common channel, while channel-by-channel convolution (DW convolution for short) has one convolution kernel responsible for one channel. DW convolution is performed in a two-dimensional plane, and the number of convolution kernels is the same as the number of channels and feature maps in the previous layer, ensuring the depth of the network and the extraction of information while reducing parameters.

[0205] (2) Point-by-point convolution. The operation of point-by-point convolution (PW convolution) is very similar to the conventional convolution operation. The size of its convolution kernel is 1x1xM, where M is the number of channels in the previous layer. Therefore, the convolution operation here will perform a weighted combination of the maps of the previous step in the depth direction to generate a new feature map. There are as many feature map outputs as there are convolution kernels. (3) Quantity calculation. The computational amount of the model is referred to as the computational amount, which is a key indicator to measure the running speed and training speed of the algorithm. The input image in the example is 3 (M) channels and 4 channels, and finally outputs 4 (N) feature maps. The length and width of the image is DF, and the convolution kernel size is (3x3)DKxDK.

[0206] The schematic diagram of the ore grade recognition system using color 3D reconstruction in this case is as follows: Figure 3 As shown in the flow chart, Figure 4 As shown in the workflow diagram Figure 5 As shown, the principle process Figure 6 shown.

[0207] The specific steps are as follows:

[0208] Step 1: Place ores of different grades into the vibrating screening and distribution system.

[0209] Step 2: Vibrating screen and belt conveyor move to drive the ore to fall off.

[0210] Step 3: Take a photo with the binocular camera and save it (the camera structure is shown in the figure below) Figure 7 shown).

[0211] Step 4: Use a surface dynamic stereoscopic vision 3D reconstruction method based on line structured light to reconstruct the photo in 3D. The specific process can be seen in step 4 above.

[0212] Step 5: Import the 3D reconstructed image into the grade recognition system and use the Mobile Net-SSA algorithm to identify the ore grade (its accuracy has been verified to be 98%).

[0213] The present invention is mainly used for ore grade identification in mining work, but the color three-dimensional reconstruction of ore can not only be used in this field, but also has an inspiring effect on other modules of mining work.

[0214] Working principle of ore grade identification system based on color 3D reconstruction

[0215] Step 1: Ore transport and distribution

[0216] 1. Ore vibration screening and conveying

[0217] The ore is placed in the vibrating screening and distribution system. The motor is turned on through the adjustment module, and the vibrating screen and belt conveyor are started. After being screened, the ore is evenly conveyed to the tail end of the belt conveyor.

[0218] The movement of the vibrating screen ensures that the ore is evenly distributed, avoiding the impact of ore accumulation on subsequent photography and identification accuracy.

[0219] 2. Ore falling off

[0220] The ore is transported to the designated location through a belt conveyor and falls off from the tail end, creating the best conditions for the binocular camera to take pictures.

[0221] Step 2: Binocular camera image acquisition

[0222] 1. Real-time photo taking

[0223] As the ore falls, the binocular camera captures real-time color images of the ore, ensuring that the surface characteristics and complete morphological information of the ore are obtained.

[0224] The camera captures fast-moving ore images by shooting at a high frame rate and transmits the data to the 3D reconstruction module in real time.

[0225] 2. Data Preprocessing

[0226] The image resolution is fixed at 256 × 256 pixels to facilitate rapid processing by the subsequent 3D reconstruction algorithm.

[0227] Step 3: 3D reconstruction

[0228] 1. Dynamic 3D reconstruction algorithm

[0229] The dynamic stereoscopic vision 3D reconstruction method based on line structured light reconstructs the images captured by the camera. This method combines global optimization with dynamic adjustment to ensure that the 3D structure of the ore can be accurately reconstructed when the ore falls at high speed.

[0230] The reconstructed image extracts multi-dimensional features by analyzing the color, shape and texture of the ore surface.

[0231] 2. Output 3D model

[0232] The reconstruction results are converted into a high-precision three-dimensional ore model, providing complete data support for subsequent grade identification.

[0233] Step 4: Taste Identification

[0234] 1. Intelligent recognition system imports images

[0235] The 3D reconstructed image is transmitted to the grade recognition module, which automatically identifies the ore grade through the Mobile Net-SSA algorithm.

[0236] Improvements of Mobile Net-SSA algorithm:

[0237] Combined with the lightweight MobileNet network, the model complexity and computational cost are significantly reduced.

[0238] Use depth-wise separable convolution to extract key features and combine it with the SSA algorithm to improve detection speed and accuracy.

[0239] 2. Ore grade determination

[0240] MobileNet extracts the features of the ore center image, combines it with the KNN model for secondary prediction, and performs weighted processing on the two results to finally output the ore grade result.

[0241] It has been verified that the system's grade recognition accuracy is as high as 98%.

[0242] Step 5: Intelligent learning and optimization

[0243] 1. Self-learning module

[0244] The system stores the recognition results in the autonomous learning module and continuously optimizes the ore grade recognition model to improve the efficiency and accuracy of the next round of data processing.

[0245] The system has a built-in knowledge base and sub-database to store images and related knowledge of different ore types for the inference engine to call in subsequent identification.

[0246] 2. Feedback and dynamic adjustment

[0247] The system continuously receives user feedback from actual operations and uses reinforcement learning to further optimize the recognition model and data processing algorithm to enhance overall adaptability.

[0248] Step 6: Result display and output

[0249] 1. Data Visualization

[0250] The system displays the ore grade identification results in real time through the interface, including 3D reconstructed images and specific grade information.

[0251] Users can view detailed identification results and obtain grade distribution statistics of the ore.

[0252] 2. Output and storage

[0253] The system supports the export function of results, which can store data to the information processing terminal for further analysis and decision-making.

[0254] This system realizes the automation of the whole process from ore transportation to grade determination by combining vibration screening, 3D reconstruction and intelligent identification. Its core technologies include dynamic 3D reconstruction algorithm and Mobile Net-SSA grade identification model, which significantly improve the speed and accuracy of ore grade identification and have broad application prospects in the fields of ore processing and resource assessment.

[0255] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0256] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A method for identifying ore grade based on color 3D reconstruction, characterized in that: The following steps are involved: After being screened by the vibrating screen, the ore is conveyed to the binocular camera through a belt conveyor and falls. The binocular camera takes real-time photos of the falling ore. The 3D point cloud of the ore surface is calculated based on the stripe center image collected by the camera and laser line generator using the dynamic stereoscopic vision 3D reconstruction method of line structured light; The reconstructed 3D image is input into the grade recognition system, and the ore grade is intelligently identified using the MobileNet-SSA algorithm. The MobileNet-SSA algorithm is combined with the KNN model for secondary prediction, and the two prediction results are weighted to output the final grade recognition result.

2. The ore grade identification method based on color three-dimensional reconstruction as claimed in claim 1, characterized in that: The three-dimensional reconstruction specifically comprises the following steps: The three-dimensional coordinates of the center point of the structured light stripes are determined by joint calibration of the binocular camera and the laser line generator; The three-dimensional point cloud of the ore surface is calculated by using the camera's perspective projection imaging model combined with the line structured light plane equation; Through relative pose estimation between adjacent frames and global optimization algorithm, the absolute pose of the binocular camera module is calculated to generate a complete three-dimensional model of the ore.

3. The ore grade identification method based on color three-dimensional reconstruction as claimed in claim 1, characterized in that: The depth-separable convolution structure of the MobileNet-SSA algorithm includes: (1) Channel-by-channel convolution: Each convolution kernel is responsible for the convolution operation of one channel, which reduces the number of parameters while ensuring the depth of information extraction; (2) Point-by-point convolution: The feature maps of channel-by-channel convolution are weighted combined in the depth direction through 1×1 convolution to generate a new feature map; (3) Computational complexity: The optimization of the model’s computational complexity makes MobileNet-SSA suitable for embedded devices and significantly reduces the model’s computational complexity while slightly reducing accuracy.

4. The ore grade identification method based on color three-dimensional reconstruction according to claim 1 is characterized in that: The MobileNet-SSA algorithm combines the advantages of Fast-RCNN and YOLO algorithms, with a detection speed that is 6 times that of Fast-RCNN and an average precision value (map) that is 11% higher than that of YOLO. MobileNet replaces the VGG16 network structure in SSA, reduces the model size and running speed requirements, and combines mask information to extract ore center image features for accurate identification.

5. An ore grade identification system based on color three-dimensional reconstruction implementing the identification method according to any one of claims 1 to 3, characterized in that: include: The regulating module is used to turn on the motor and adjust the speed regulator to make the vibrating screen and belt conveyor move; The camera module is used for the ore to fall to the tail end of the belt conveyor. The camera takes pictures and sends the images to the 3D reconstruction system. A three-dimensional reconstruction module, used for performing three-dimensional reconstruction of the input image; The grade recognition module is used to transmit the 3D reconstructed image to the intelligent recognition system for grade recognition based on the MobileNet-SSA algorithm.

6. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the ore grade identification method based on color three-dimensional reconstruction as described in any one of claims 1 to 3.

7. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the ore grade identification method based on color three-dimensional reconstruction as claimed in any one of claims 1 to 3.

8. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the ore grade identification system based on color three-dimensional reconstruction as described in claim 4.

Images