A tire qualification detection method based on a multi-view camera

Abstract

The application discloses a kind of based on multi-view camera's tire qualified detection method, comprising: multi-view camera calibration, including the internal parameter of camera, external parameter and the parameter between camera;Tire image pre-processing, including the image registration fusion of multi-view camera acquisition, then image is grayed, enhanced and multi-angle Canny operator edge detection processing, extract for detecting tread image;Tire image target detection includes size detection and pattern recognition, size detection will tread image symmetry two points with parabola differentiation and integration principle to calculate arc length and set value comparison, pattern recognition is combined by the LBP algorithm of elliptical area and hash algorithm, to effectively compare target pattern.Compared with prior art, the present application can better extract the tread image, improve the accuracy and robustness, improve the speed and detection accuracy of detection, suitable for industrial tire production detection.

Description

Technical Field

[0001] This invention relates to a tire quality inspection method based on a multi-view camera, belonging to the technical field of industrial inspection. Background Technology

[0002] Traditional tire quality inspection relies on manual labor, which is labor-intensive, highly susceptible to subjective bias, and inefficient. While X-ray-based tire defect systems have emerged abroad with advancements in industrial inspection technology, their algorithms are difficult to apply to domestic tire manufacturers due to differences in defect assessment standards. The core process still relies on manual observation and judgment of X-ray images, failing to address the issue of subjective bias. The rapid development of machine vision has effectively solved the drawbacks of manual inspection.

[0003] Given the uniform color and limited feature points of tire treads, effective registration and fusion of images captured by multi-view cameras present significant challenges. In extracting effective patterns and contours from tread images, the single-angle Canny operator has limitations and cannot extract sufficient information for comparison. Currently, Gamma image enhancement and edge denoising techniques have achieved considerable success in their respective areas. Effectively combining these two techniques can significantly improve image information extraction and target recognition processes. Summary of the Invention

[0004] Purpose of the invention: This invention designs a tire quality inspection method based on a multi-view camera. The multi-view camera and multi-angle Canny edge detection are used to better extract the tire tread image. The improved algorithms for the two target detections also improve accuracy and robustness. This detection method improves the detection speed and accuracy, and is suitable for industrial tire production inspection.

[0005] Technical solution: A tire quality inspection method based on multi-view cameras, including multi-view camera calibration, tire image preprocessing, and target detection of tire images;

[0006] Step 1: Calibration of multi-camera: The calibration content includes the internal parameters, external parameters, and parameters between the cameras. The internal parameters include the magnification factor of the camera in the horizontal and vertical directions, the coupling magnification factor in the horizontal and vertical directions, the horizontal and vertical coordinates of the captured image, and the coefficients of tangential and radial distortion. The external parameters include the rotation matrix and translation matrix between the cameras.

[0007] Step 2: Preprocessing of tire images: Multi-angle image registration and image fusion using the direct averaging fusion method are performed on the images acquired by the multi-camera. Then, Gamma transformation enhancement and multi-angle Canny operator edge detection are performed on the images to extract the tire tread image for detection.

[0008] Step 3: Target detection in tire image: This includes size detection and pattern recognition. Size detection calculates the arc length of two symmetrical points in the tire tread image using the principles of parabolic differentiation and integration, and compares it with a set value. Pattern recognition uses a combination of the LBP algorithm and hash algorithm for elliptical regions to effectively compare the target pattern.

[0009] Furthermore, the calibration step in step 1 specifically includes:

[0010] (1) Install and fix the four-channel camera. The cameras are set in sequence as camera 1, camera 2, camera 3, and camera 4. Place the calibration plate in advance, select camera 1, and adjust the focus.

[0011] (2) Use camera 1 to take a straight picture of the calibration board to obtain an image of the target array n×m;

[0012] (3) Establish the camera's internal parameter model, which has 5 parameters. The model is as follows:

[0013]

[0014] Among them, f x f is the horizontal magnification factor of the camera. y f is the vertical magnification factor of the camera. z c represents the coupling amplification factor in the horizontal and vertical directions. x and c y The x and y coordinates of the center of the image captured by the camera, (X c Y c Z c (u, v) is the world coordinate system, (u, v) is the pixel coordinate system, and f / Z c These are the parameters for perspective projection;

[0015] (4) Establish the extrinsic model of the camera, which has two parameters, as follows:

[0016]

[0017] Where R represents the rotation matrix, t is the translation matrix, (X w Y w Z w (X) is the world coordinate system. c Y c Zc () is the camera coordinate system;

[0018] (5) Combining the five internal parameter models and the two external parameter models, the five internal parameters and the two external parameters are obtained by using the properties of rotation matrices and real symmetric matrices.

[0019] (6) Based on the five intrinsic parameters obtained, calculate the coordinates of the image point on the focal length-normalized imaging plane, and based on the radial and tangential distortion models, calculate the radial and tangential distortion coefficients to establish a distortion model:

[0020]

[0021] Where (u, v) are the actual coordinates of the image, and r is the distance from the image point (u′, v′) to the image center (0, 0). k1 and k2 are the second-order and fourth-order radial distortion coefficients, respectively, and p1 and p2 are the second-order and fourth-order tangential distortion coefficients, respectively.

[0022] Furthermore, the specific process for solving the parameters between the cameras in step 1 includes:

[0023] (1) Consider a point P in space, with coordinates (X, Y, Z) in the world coordinate system, and the points imaged in the camera coordinate systems of camera 1 and camera 2 as (x1, y1) and (x2, y2) respectively. f represents the effective focal length, and L represents the distance between the optical centers of the two cameras. The relationship between the two coordinate systems is as follows:

[0024]

[0025] Thus, the relationship between camera 1 and camera 2 is as follows:

[0026]

[0027] Where Z(x1-x2)=Lf, the parameters between cameras 1 and 2 are obtained;

[0028] (2) The points imaged in the camera coordinate systems of cameras 2 and 3 are (x2, y2) and (x3, y3) respectively. The relationship between the two coordinate systems is:

[0029]

[0030] Thus, the relationship between camera 2 and camera 3 is as follows:

[0031]

[0032] Where Z(x2-x3)=Lf, the parameters between cameras 2 and 3 are obtained.

[0033] (3) The points imaged in the camera coordinate systems of camera 3 and camera 4 are (x3, y3) and (x4, y4) respectively. The relationship between the two coordinate systems is:

[0034]

[0035] Thus, the relationship between camera 3 and camera 4 is as follows:

[0036]

[0037] Where Z(x3-x4)=Lf, the parameters between cameras 3 and 4 are obtained.

[0038] Furthermore, the specific operations in step 2 include:

[0039] (1) The multi-camera is set to the same frequency to shoot, and the captured images are transmitted synchronously using the frame synchronization method. The frame synchronization selects the falling edge signal as the start of a frame image data transmission. This sequence is called the "frame header". The host computer determines the start of image transmission by verifying the "frame header" data. If the amount of data received before the next "frame header" arrives meets the data amount corresponding to a frame image, the transmission is considered successful. Otherwise, it will be considered a failure and the transmission will be restarted.

[0040] (2) The process of stitching together multiple images taken at the same frequency is to use the method of calculating the transform domain of the image to be registered using a calibration plate for image registration, and to use the hom_mat2d operator in Halcon for calculation.

[0041] (3) By manually selecting the four points in the calibration plate image, determine the homogeneous projective transformation matrix HomMat2D corresponding to the optimal solution that satisfies the following equation, as shown below:

[0042]

[0043] Among them, (P) x P y ), (Q x Q y These are points in the manually selected base image and target image, respectively.

[0044] (4) Rotate the target image to make it parallel to the corresponding reference image plane. Rotate the image according to the calculated HomMat2D matrix, then crop part of the image, remove the unnecessary background area, and finally calculate the initial translation amount in the image according to the calculated projection transformation matrix to register the target image with the base image.

[0045] (5) After image registration, the registered images are fused using the direct averaging fusion method. The gray values ​​of the overlapping areas of the target image and the base image are all set to the average value of the overlapping areas. The specific formula is as follows:

[0046]

[0047] Where I1 represents the base image, I2 represents the target image, and I represents the fused image.

[0048] Furthermore, step 2 involves performing Gamma transformation enhancement and multi-angle Canny operator edge detection on the image to extract the image used for tire tread detection, specifically including:

[0049] (1) The Gamma correction algorithm in the nonlinear enhancement algorithm is used to enhance the image. The Gamma transformation is used to optimize the tire image and adjust the local gray level in the image to the best effect, so that the pixel gray level values ​​between the tire edge and the background in the image are significantly different, which can effectively distinguish the tire and background pixels.

[0050] (2) Multi-angle Canny operator edge detection: First, a two-dimensional Gaussian function is convolved with the enhanced image to filter and denoise, resulting in a smooth image. The two-dimensional Gaussian function G(x,y) is:

[0051]

[0052] The formula for the convolution of a Gaussian function with an enhanced image is:

[0053] R(x,y)=G(x,y)*f(x,y) (11)

[0054] In the formula: σ is the standard deviation of the Gaussian function, and its magnitude can change the smoothness of the image; R(x,y) is the image after the Gaussian function is convolved with the input image; f(x,y) is the input image; * represents the convolution operation, which converts the Gaussian function into a Gaussian filter template;

[0055] (3) Select a 3×3 Sobel operator, and convolve the gradient templates of the eight edge directions (0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°) with the filtered image to obtain the gradient components in the eight directions, M. 0° M 45° M 90° M 135° M 180° M 225° M 270° M 315° Calculate the absolute value of the gradient magnitude M:

[0056]

[0057] The gradient components in the eight directions are synthesized into a total degree of M in the horizontal and vertical directions. x M y The following formula can be seen:

[0058]

[0059]

[0060] Its gradient direction for:

[0061]

[0062] After obtaining a histogram based on all pixel values ​​in the image, high and low thresholds are obtained. The high threshold is used to divide strong edges, and the low threshold is used to divide weak edges.

[0063] 8. Further, the specific steps for size detection in step 3 are as follows:

[0064] Two arbitrary points are selected on the edge tire image obtained by Canny operator edge detection to calculate the arc length. The arc between the two points can be approximated as a parabola. A rectangular coordinate system is established with the first point A(X1, Y1) as the origin, and the coordinates of the two points are A(0, 0) and B(X2, Y2). The parabola model is shown below:

[0065] y = ax 2 +bx (16)

[0066]

[0067] Y²x - X²y = 0 (18)

[0068] Equation (16) is the equation of line AB. This model utilizes the lowest point of the parabola as... By considering the relationship between the distance to line AB and the sag m, we can solve for the equation of the parabola, whose model is as follows:

[0069]

[0070] Combining equations (15) and (17), the unknown parameters a and b can be solved. Then, the arc length L between the two points can be calculated using an integral model, which is as follows:

[0071]

[0072] By combining the arc length L with the angle θ between the two points, the radius R can be calculated and compared with the set value to determine whether the tire size meets the standard.

[0073] 9. Further, the specific steps for pattern detection in step 3 are as follows:

[0074] (1) Pattern recognition establishes an elliptical LBP operator model to improve the adaptability of image size. Its mathematical model is as follows:

[0075]

[0076] Where a is the semi-major axis of the neighborhood, b is the semi-minor axis of the neighborhood, N is the number of equally spaced pixels selected on the ellipse, and the coordinates of the center point are (x, y). c y c When ), the coordinates of the i-th pixel adjacent to the center point are (x... i y i ), where i∈N;

[0077] The mathematical solution process for LBP codes is modeled as follows:

[0078]

[0079]

[0080] Where N is the number of edge pixels, g i Let g be the pixel value of the i-th pixel adjacent to the center point. c The pixel value at the center point;

[0081] To maintain consistency before and after image rotation, the LBP codes obtained by shifting the image to the right by 's' bits each time are compared to obtain the largest LBP code:

[0082] LBP=max{R(LBP,s)|s=0,1,Λ,N-1} (22)

[0083] Wherein, R(LBP, s) is the result obtained by shifting the LBP code with N sampling points to the right by s bits. The largest LBP code can be obtained by comparison.

[0084] (2) Extract image patterns from LBP images in rotation-invariant mode, and calculate the Hamming distance between the two images using a hash algorithm. The model is as follows:

[0085]

[0086] Where x and y are the hash codes of the two target images, x[i] and y[i] are the i-th elements of x and y respectively, and d(x, y) is the Hamming distance between the two images;

[0087] The similarity between two images of size m×n and Hamming distance d is:

[0088]

[0089] When the similarity (sim) value is above 95%, the tire tread pattern is considered to have passed the inspection.

[0090] Beneficial effects:

[0091] The multi-view camera used in this invention can capture images of the tire from multiple angles, making the detection results more stable and reliable. Multi-angle Canny edge detection is employed to better extract the tire's contour and tread pattern. The size detection in the target recognition process uses arc length calculus calculation, resulting in more accurate tire size detection. Tread pattern detection combines LBP and hash algorithms to improve detection accuracy. In summary, this invention meets the requirements for tire detection. Attached Figure Description

[0092] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0093] Figure 2 This is a flowchart of the multi-view camera calibration process of the present invention;

[0094] Figure 3 A flowchart for image preprocessing;

[0095] Figure 4 This is a 3×3 template diagram for the multi-angle Sobel operator. Detailed Implementation

[0096] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0097] like Figure 1 As shown, this invention discloses a tire quality inspection method based on a multi-view camera, comprising the following steps:

[0098] Step 1: Calibration of the multi-view camera.

[0099] Step 2: Preprocessing of tire images.

[0100] Step 3: Target detection in tire images.

[0101] In the multi-camera calibration step of step 1 above, the camera's internal parameters, external parameters, and inter-camera parameters are obtained based on the captured calibration board; the internal parameters include the camera's horizontal magnification factor f. x Vertical magnification factor fy The coupling amplification factor f in the horizontal and vertical directions z The horizontal coordinate c of the center of the image captured by the camera x and the vertical coordinate c y , as well as tangential and radial distortion coefficients; external parameters include rotation matrix R and translation matrix T.

[0102] The calibration steps specifically include:

[0103] (1) Install and fix the four-channel camera. The cameras are set in sequence as camera 1, camera 2, camera 3, and camera 4. Place the calibration plate in advance, select camera 1, and adjust the focus.

[0104] (2) Use camera 1 to shoot the calibration board directly to obtain the image of the target array n×m.

[0105] (3) Establish the camera's internal parameter model, which has 5 parameters. The model is as follows:

[0106]

[0107] Among them, f x f is the horizontal magnification factor of the camera. y f is the vertical magnification factor of the camera. z c represents the coupling amplification factor in the horizontal and vertical directions. x and c y The x and y coordinates of the center of the image captured by the camera, (X c Y c Z c (u, v) is the world coordinate system, (u, v) is the pixel coordinate system, and f / Z c These are the parameters for perspective projection;

[0108] (4) Establish the extrinsic model of the camera, which has two parameters, as follows:

[0109]

[0110] Where R represents the rotation matrix, t is the translation matrix, (X w Y w Z w (X) is the world coordinate system. c Y c Z c () is the camera coordinate system;

[0111] (5) Combining equations (1) and (2), by utilizing the properties of rotation matrices and real symmetric matrices, we can find five internal parameters and two external parameters;

[0112] (6) Based on the five intrinsic parameters obtained, calculate the coordinates of the image point on the focal length-normalized imaging plane, and based on the radial and tangential distortion models, calculate the radial and tangential distortion coefficients to establish a distortion model:

[0113]

[0114] Where (u, v) are the actual coordinates of the image, and r is the distance from the image point (u′, v′) to the image center (0, 0). k1 and k2 are the second-order and fourth-order radial distortion coefficients, respectively, and p1 and p2 are the second-order and fourth-order tangential distortion coefficients, respectively.

[0115] like Figure 2 As shown, the parameters between the cameras are solved in the following ways:

[0116] (1) Consider a point P in space, with coordinates (X, Y, Z) in the world coordinate system, and the points imaged in the camera coordinate systems of camera 1 and camera 2 as (x1, y1) and (x2, y2) respectively. f represents the effective focal length, and L represents the distance between the optical centers of the two cameras. The relationship between the two coordinate systems is as follows:

[0117]

[0118] Thus, the relationship between camera 1 and camera 2 is as follows:

[0119]

[0120] Where Z(x1-x2)=Lf, the parameters between cameras 1 and 2 are obtained.

[0121] (2) The points imaged in the camera coordinate systems of cameras 2 and 3 are (x2, y2) and (x3, y3) respectively. The relationship between the two coordinate systems is:

[0122]

[0123] Thus, the relationship between camera 2 and camera 3 is as follows:

[0124]

[0125] Where Z(x2-x3)=Lf, the parameters between cameras 2 and 3 are obtained.

[0126] (3) The points imaged in the camera coordinate systems of camera 3 and camera 4 are (x3, y3) and (x4, y4) respectively. The relationship between the two coordinate systems is:

[0127]

[0128] Thus, the relationship between camera 3 and camera 4 is as follows:

[0129]

[0130] Where Z(x3-x4)=Lf, the parameters between cameras 3 and 4 are obtained.

[0131] like Figure 3 As shown, the image preprocessing steps in step 2 specifically include:

[0132] (1) When multiple cameras are set to the same frequency, they shoot and transmit the captured images synchronously using the frame synchronization method. The frame synchronization selects the falling edge signal as the start of a frame of image data transmission. This sequence is called the "frame header". The host computer determines the start of image transmission by verifying the "frame header" data. If the amount of data received before the next "frame header" arrives meets the data amount corresponding to a frame of image, the transmission is considered successful. Otherwise, it is considered a failure and the transmission is restarted.

[0133] (2) Image registration is the process of stitching together multiple images taken at the same frequency. The method of calculating the transform domain of the image to be registered using a calibration plate is selected for image registration. Since the tire surface has few feature points and a single color, the hom_mat2d operator in Halcon is mainly used for calculation.

[0134] (3) By manually selecting the four points in the calibration plate image, determine the homogeneous projective transformation matrix HomMat2D corresponding to the optimal solution that satisfies the following equation, as shown below:

[0135]

[0136] Among them, (P) x P y ), (Q x Q y ( ) are points in the manually selected base image and target image, respectively.

[0137] (4) First, rotate the target image to make it parallel to the corresponding reference image plane. Rotate the image according to the calculated HomMat2D matrix. Then, crop a portion of the image, removing unnecessary areas such as the background. Finally, calculate the initial translation amount in the image according to the calculated projection transformation matrix to register the target image with the base image.

[0138] (5) After image registration, the image stitching process is basically completed. However, it can be seen that there is a clear stitching seam between the two stitched images. Therefore, the registered images need to be fused. The direct averaging fusion method is adopted, and the gray values ​​of the overlapping areas of the target image and the base image are all set to the average value of the overlapping area. This can be expressed by the following formula:

[0139]

[0140] Where I1 represents the base image, I2 represents the target image, and I represents the fused image.

[0141] (6) The Gamma correction algorithm in the nonlinear enhancement algorithm is used to enhance the image. The tire image is optimized by using Gamma transformation to adjust the local gray level in the image to the best effect, so that the pixel gray level values ​​between the tire edge and the background in the image are significantly different, which can effectively distinguish the tire surface and the background pixels.

[0142] (7) Multi-angle Canny edge detection: First, a two-dimensional Gaussian function is convolved with the enhanced image to filter and denoise, resulting in a smooth image. The two-dimensional Gaussian function G(x, y) is:

[0143]

[0144] The formula for the convolution of a Gaussian function with an enhanced image is:

[0145] R(x,y)=G(x,y)*f(x,y) (11)

[0146] In the formula: σ is the standard deviation of the Gaussian function, and its magnitude can change the smoothness of the image; R(x,y) is the image after the Gaussian function is convolved with the input image; f(x,y) is the input image; * represents the convolution operation, which converts the Gaussian function into a Gaussian filter template, which can improve the computational efficiency.

[0147] (8) Figure 4 As shown, a 3×3 Sobel operator is selected, and the gradient templates of the eight edge directions (0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°) are convolved with the filtered image to obtain the gradient components in the eight directions, M. 0° M 45° M 90° M 135° M 180° M 225° M 270° M 315° Calculate the absolute value of the gradient magnitude M:

[0148]

[0149] The gradient components in the eight directions are synthesized into a total degree of M in the horizontal and vertical directions. x M y As can be seen from equations (13) and (14) below.

[0150]

[0151]

[0152] Its gradient direction for:

[0153]

[0154] After obtaining a histogram based on all pixel values ​​in the image, high and low thresholds are obtained. The high threshold is used to divide strong edges, and the low threshold is used to divide weak edges.

[0155] The target detection in step 3 specifically includes:

[0156] (1) Take any two points on the edge tire image obtained by the Canny operator to calculate the arc length. The arc between the two points can be approximated as a parabola. Establish a rectangular coordinate system with the first point A(X1, Y1) as the origin. The coordinates of the two points are A(0, 0) and B(X2, Y2). The parabola model is shown below:

[0157] v = ax 2 +bx (16)

[0158]

[0159] Y²x - X²y = 0 (18)

[0160] Equation (16) is the equation of line AB. This model utilizes the lowest point of the parabola as... By considering the relationship between the distance to line AB and the sag m, we can solve for the equation of the parabola, whose model is as follows:

[0161]

[0162] Combining equations (15) and (17), the unknown parameters a and b can be solved. Then, the arc length L between the two points can be calculated using an integral model, which is as follows:

[0163]

[0164] By combining the arc length L and the angle θ between the two points, the radius R can be solved and compared with the set value to determine whether the tire size meets the standard.

[0165] (2) Pattern recognition establishes an elliptical LBP operator model to improve the adaptability of image size. Its mathematical model is shown below:

[0166]

[0167] Where a is the semi-major axis of the neighborhood, b is the semi-minor axis of the neighborhood, N is the number of equally spaced pixels selected on the ellipse, and the coordinates of the center point are (x, y). c y c When ), the coordinates of the i-th pixel adjacent to the center point are (x... i y i ), where i∈N.

[0168] The mathematical solution process for LBP codes is modeled as follows:

[0169]

[0170]

[0171] Where N is the number of edge pixels, g i Let g be the pixel value of the i-th pixel adjacent to the center point. c The pixel value at the center point;

[0172] To maintain consistency before and after image rotation, the LBP codes obtained by shifting the image to the right by 's' bits each time are compared to obtain the largest LBP code:

[0173] LBP=max{R(LBP,s)|s=0,1,Λ,N-1} (24)

[0174] R(LBP, s) is the result obtained by shifting the LBP code with N sampling points to the right by s bits. By comparison, the largest LBP code can be obtained, which effectively overcomes the image rotation problem and ensures the consistency of the image before and after rotation.

[0175] (3) Extract image patterns from LBP images in rotation-invariant mode, and calculate the Hamming distance between the two images using a hash algorithm. The model is as follows:

[0176]

[0177] Where x and y are the hash codes of the two target images, x[i] and y[i] are the i-th elements of x and y respectively, and d(x, y) is the Hamming distance between the two images.

[0178] The similarity between two images of size m×n and Hamming distance d is:

[0179]

[0180] When the similarity (sim) value is above 95%, the tire tread pattern is considered to have passed the inspection.

[0181] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent transformations or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A multi-camera based tire qualification method, characterized in that, It comprises the following steps: Step 1: Calibration of multi-camera: the calibration content is the internal parameters of the camera, the external parameters and the parameters between the cameras, the internal parameters include the magnification coefficients of the camera in horizontal and vertical directions, the coupled magnification coefficients in horizontal and vertical directions, the horizontal and vertical coordinates of the captured image and the coefficients of tangential distortion and radial distortion, the external parameters include the rotation matrix and translation matrix between the cameras; Step 2: Pretreatment of tire image: multi-angle image registration is performed on the images collected by the multi-camera, and image fusion is performed using the direct average fusion method, then Gamma transformation enhancement and multi-angle Canny operator edge detection processing are performed on the image to extract the tire image for detection; Step 3: Target detection of tire image: including size detection and pattern recognition, the size detection calculates the arc length by using the principle of parabolic differentiation and integration of the two symmetrical points of the tire image, and compares it with the set value, the pattern recognition effectively compares the target pattern by combining the LBP algorithm and the hash algorithm of the elliptical region; The specific steps of size detection in step 3 are as follows: Any two points on the edge tire image obtained by Canny operator edge detection are used to calculate the arc length, the circular arc between the two points is approximated as a parabola, and a rectangular coordinate system is established with the first point A(X1, Y1) as the origin, the coordinates of the two points are A(0, 0) and B(X2, Y2), and the parabola model is as follows: y = ax 2 + bx (16) Y2x-X2y=0 (18) where equation (16) is the equation of the straight line AB, and the model is as follows by using the parabola minimum point as the relationship between the distance to the straight line AB and the sag m, and the parabola equation is solved, and the model is as follows: The unknown parameters a and b are solved by combining formula (19) and formula (17), and then the arc length L between the two points is calculated by using the integral model, and the model is as follows: The radius R is solved by combining the arc length L and the angle θ between the two points, and compared with the set value to determine whether the tire size meets the standard; The specific steps of pattern detection in step 3 are as follows: (1) The LBP operator model of the elliptical pattern recognition is established to improve the adaptability of the image size, and the mathematical model is as follows: Wherein, a is the half length of the semi-neighbor, b is the half short axis of the semi-neighbor, N is the number of equally spaced pixel points selected on the ellipse, the coordinates of the center point are (x c ,y c ), the coordinates of the i-th pixel point adjacent to the center point are (x i ,y i ), wherein i∈N; The mathematical solution process model of LBP code is as follows: wherein N is the number of edge pixels, g i is the pixel value of the i-th pixel adjacent to the center point, g c is the pixel value of the center point. On the problem of keeping the image consistent before and after rotation, the LBP codes obtained by moving s bits to the right each time are compared with each other to obtain the maximum LBP code: LBP=max{R(LBP,s)|s=0,1,…,N-1} (24) Where R(LBP, s) is the result obtained by moving s bits to the right of the LBP code with a sample point number of N, and the maximum LBP code can be obtained by comparison; (2) The image pattern is extracted from the LBP image in the rotation invariant mode, and the Hamming distance between the two images is calculated by the hash algorithm, and the model is as follows: Where x and y are the hash codes of the two target images, x[i] and y[i] are the i-th elements in x and y respectively, and d(x, y) is the Hamming distance between the two images; The similarity of two images with size m*n and Hamming distance d is: When the value of similarity sim is above 95%, it is considered that the tire pattern detection is passed.

2. The multi-camera based tire pass / fail detection method of claim 1, wherein, The calibration step of step 1 specifically comprises: (1) Install and fix the four-channel camera, and set the camera in order as camera 1, camera 2, camera 3 and camera 4. Place the calibration board in advance, select camera 1, and adjust the focal length; (2) Use camera 1 to directly shoot the calibration board to obtain the target array n x m image; (3) Establish the camera internal parameter model, which has five parameters, and the model is as follows: wherein f x is a horizontal magnification factor of the camera, f y is a vertical magnification factor of the camera, f z is a coupling magnification factor of the horizontal and vertical directions, c x and c y are the horizontal and vertical coordinates of the image center of the camera, (X c , Y c , Z c ) is a world coordinate system, (u, v) is a pixel coordinate system, and f / Z c is a parameter of the perspective projection. (4) Establish the camera external parameter model, which has two parameters, and the model is as follows: wherein R represents a rotation matrix, t is a translation matrix, (X w ,Y w ,Z w ) is a world coordinate system, and (X c ,Y c ,Z c ) is a camera coordinate system; (5) Combine the five internal parameter models and two external parameter models, and use the properties of rotation matrix and real symmetric matrix to solve the five internal parameters and two external parameters; (6) According to the five internal parameters obtained, the imaging point coordinates of the image point on the focal length normalized imaging plane are solved, and the radial distortion and tangential distortion coefficients are solved based on the radial distortion and tangential distortion model to establish the distortion model: where (u, v) is the actual image coordinate, r is the distance from the image point (u', v') to the image center (0, 0), k1, k2 are the second and fourth order radial distortion coefficients, respectively, and p1, p2 are the second and fourth order tangential distortion coefficients, respectively.

3. The multi-camera based tire pass / fail detection system of claim 2, wherein, The specific solving process of the parameters between the cameras in step 1 includes: (1) Considering a space point P, the coordinates of which in the world coordinate system are (X, Y, Z), the imaging points of which in the camera coordinate systems of camera 1 and camera 2 are (x1, y1) and (x2, y2) respectively, f represents the effective focal length, L represents the distance between the optical centers of the two cameras, and the relationship between the two coordinate systems is: Thus, the relationship from camera 1 to camera 2 is: Wherein, Z(x1-x2)=Lf, the parameters between camera 1 and camera 2 are solved; (2) The imaging points of which in the camera coordinate systems of camera 2 and camera 3 are (x2, y2) and (x3, y3) respectively, and the relationship between the two coordinate systems is: Thus, the relationship from camera 2 to camera 3 is: Wherein, Z(x2-x3)=Lf, the parameters between camera 2 and camera 3 are solved; (3) The imaging points of which in the camera coordinate systems of camera 3 and camera 4 are (x3, y3) and (x4, y4) respectively, and the relationship between the two coordinate systems is: Thus, the relationship from camera 3 to camera 4 is: Wherein, Z(x3-x4)=Lf, the parameters between camera 3 and camera 4 are solved.

4. The multi-camera based tire pass / fail detection method of claim 2, wherein, The specific operation in step 2 includes: (1) The multi-view camera is set to shoot at the same frequency, and the image synchronization transmission is realized by using frame synchronization method, the frame synchronization is realized by selecting the falling edge signal as the start of a frame of image data transmission, which is called "frame header", the host computer determines the start of image transmission by checking the "frame header" data, before the next "frame header" arrives, the amount of data accepted meets the data amount corresponding to a frame of image, then the transmission is considered successful, otherwise it is considered failed, and the transmission is re-performed; (2) The process of splicing multiple images shot at the same frequency, using the method of calculating the transformation domain of the to-be-registered image to register the images, and using the hom_mat2d operator in Halcon to calculate; (3) The four points in the calibration board image corresponding to the optimal solution are determined by manual selection, and the homogeneous projective transformation matrix HomMat2D satisfying the following equation is obtained: wherein (P x , P y ) and (Q x , Q y ) are manually selected points of the base image and the target image, respectively; (4) Rotating the target image to be parallel to the corresponding reference image plane, rotating the image according to the calculated HomMat2D matrix, then cutting the partial region of the image, cutting the region not needed in the background, and finally calculating the initial translation in the image according to the calculated projection transformation matrix, so that the target image is registered with the base image; (5) After image registration, the registered picture is fused, and the direct average fusion method is used, that is, the gray value of the overlapping region of the target image and the base image is set to the average value of the overlapping region, and the specific formula is: Where I1 represents the base image, I2 represents the target image, and I represents the fused image.

5. The multi-camera based tire pass / fail inspection method of claim 1, wherein, The Gamma transformation enhancement and multi-angle Canny operator edge detection processing of the image in step 2 are used to extract the tire image for detection, which specifically includes: (1) The Gamma correction algorithm in the nonlinear enhancement algorithm is used to enhance the image, and the tire image is optimized by using Gamma transformation, so that the local gray scale in the image is adjusted to the best effect, the pixel gray scale difference between the tire edge and the background in the image is large, and the tire and the background pixels can be well distinguished; (2) The multi-angle Canny operator edge detection first uses a two-dimensional Gaussian function to perform convolution operation with the enhanced image for filtering and denoising to obtain a smooth image, and the two-dimensional Gaussian function G(x, y) is: The convolution formula of the Gaussian function and the enhanced image is: R(x, y) = G(x, y) * f(x, y) (11) In the formula, sigma is the standard deviation of the Gaussian function, and the change of the size changes the smoothing degree of the image; R(x, y) is the image after the convolution operation of the Gaussian function and the input image; f(x, y) is the input image; * represents convolution operation, which converts the Gaussian function into a Gaussian filter template; (3) Select a 3x3 Sobel operator, and perform convolution operation on the filtered image with the gradient templates of eight edge directions of 0°, 45°, 90°, 135°, 180°, 225°, 270° and 315° respectively to obtain eight directional gradient components M 0° , 45° , 90° , 135° , 180° , 225° , 270° , 315° Calculate the absolute value of the gradient amplitude M: The gradient components in the eight directions are synthesized into two directions, horizontal and vertical, with total degrees M x , M y , the following equation is seen: the gradient direction thereof is: The high threshold is used to divide the strong edge, and the low threshold is used to divide the weak edge.

Images